1 /* 2 * Copyright (c) 1998, 2018, Oracle and/or its affiliates. All rights reserved. 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. 4 * 5 * This code is free software; you can redistribute it and/or modify it 6 * under the terms of the GNU General Public License version 2 only, as 7 * published by the Free Software Foundation. 8 * 9 * This code is distributed in the hope that it will be useful, but WITHOUT 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License 12 * version 2 for more details (a copy is included in the LICENSE file that 13 * accompanied this code). 14 * 15 * You should have received a copy of the GNU General Public License version 16 * 2 along with this work; if not, write to the Free Software Foundation, 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. 18 * 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA 20 * or visit www.oracle.com if you need additional information or have any 21 * questions. 22 * 23 */ 24 25 #include "precompiled.hpp" 26 #include "ci/ciMethodData.hpp" 27 #include "compiler/compileLog.hpp" 28 #include "gc/shared/barrierSet.hpp" 29 #include "gc/shared/c2/barrierSetC2.hpp" 30 #include "libadt/vectset.hpp" 31 #include "memory/allocation.inline.hpp" 32 #include "memory/resourceArea.hpp" 33 #include "opto/addnode.hpp" 34 #include "opto/callnode.hpp" 35 #include "opto/connode.hpp" 36 #include "opto/convertnode.hpp" 37 #include "opto/divnode.hpp" 38 #include "opto/idealGraphPrinter.hpp" 39 #include "opto/loopnode.hpp" 40 #include "opto/mulnode.hpp" 41 #include "opto/rootnode.hpp" 42 #include "opto/superword.hpp" 43 44 //============================================================================= 45 //------------------------------is_loop_iv------------------------------------- 46 // Determine if a node is Counted loop induction variable. 47 // The method is declared in node.hpp. 48 const Node* Node::is_loop_iv() const { 49 if (this->is_Phi() && !this->as_Phi()->is_copy() && 50 this->as_Phi()->region()->is_CountedLoop() && 51 this->as_Phi()->region()->as_CountedLoop()->phi() == this) { 52 return this; 53 } else { 54 return NULL; 55 } 56 } 57 58 //============================================================================= 59 //------------------------------dump_spec-------------------------------------- 60 // Dump special per-node info 61 #ifndef PRODUCT 62 void LoopNode::dump_spec(outputStream *st) const { 63 if (is_inner_loop()) st->print( "inner " ); 64 if (is_partial_peel_loop()) st->print( "partial_peel " ); 65 if (partial_peel_has_failed()) st->print( "partial_peel_failed " ); 66 } 67 #endif 68 69 //------------------------------is_valid_counted_loop------------------------- 70 bool LoopNode::is_valid_counted_loop() const { 71 if (is_CountedLoop()) { 72 CountedLoopNode* l = as_CountedLoop(); 73 CountedLoopEndNode* le = l->loopexit_or_null(); 74 if (le != NULL && 75 le->proj_out_or_null(1 /* true */) == l->in(LoopNode::LoopBackControl)) { 76 Node* phi = l->phi(); 77 Node* exit = le->proj_out_or_null(0 /* false */); 78 if (exit != NULL && exit->Opcode() == Op_IfFalse && 79 phi != NULL && phi->is_Phi() && 80 phi->in(LoopNode::LoopBackControl) == l->incr() && 81 le->loopnode() == l && le->stride_is_con()) { 82 return true; 83 } 84 } 85 } 86 return false; 87 } 88 89 //------------------------------get_early_ctrl--------------------------------- 90 // Compute earliest legal control 91 Node *PhaseIdealLoop::get_early_ctrl( Node *n ) { 92 assert( !n->is_Phi() && !n->is_CFG(), "this code only handles data nodes" ); 93 uint i; 94 Node *early; 95 if (n->in(0) && !n->is_expensive()) { 96 early = n->in(0); 97 if (!early->is_CFG()) // Might be a non-CFG multi-def 98 early = get_ctrl(early); // So treat input as a straight data input 99 i = 1; 100 } else { 101 early = get_ctrl(n->in(1)); 102 i = 2; 103 } 104 uint e_d = dom_depth(early); 105 assert( early, "" ); 106 for (; i < n->req(); i++) { 107 Node *cin = get_ctrl(n->in(i)); 108 assert( cin, "" ); 109 // Keep deepest dominator depth 110 uint c_d = dom_depth(cin); 111 if (c_d > e_d) { // Deeper guy? 112 early = cin; // Keep deepest found so far 113 e_d = c_d; 114 } else if (c_d == e_d && // Same depth? 115 early != cin) { // If not equal, must use slower algorithm 116 // If same depth but not equal, one _must_ dominate the other 117 // and we want the deeper (i.e., dominated) guy. 118 Node *n1 = early; 119 Node *n2 = cin; 120 while (1) { 121 n1 = idom(n1); // Walk up until break cycle 122 n2 = idom(n2); 123 if (n1 == cin || // Walked early up to cin 124 dom_depth(n2) < c_d) 125 break; // early is deeper; keep him 126 if (n2 == early || // Walked cin up to early 127 dom_depth(n1) < c_d) { 128 early = cin; // cin is deeper; keep him 129 break; 130 } 131 } 132 e_d = dom_depth(early); // Reset depth register cache 133 } 134 } 135 136 // Return earliest legal location 137 assert(early == find_non_split_ctrl(early), "unexpected early control"); 138 139 if (n->is_expensive() && !_verify_only && !_verify_me) { 140 assert(n->in(0), "should have control input"); 141 early = get_early_ctrl_for_expensive(n, early); 142 } 143 144 return early; 145 } 146 147 //------------------------------get_early_ctrl_for_expensive--------------------------------- 148 // Move node up the dominator tree as high as legal while still beneficial 149 Node *PhaseIdealLoop::get_early_ctrl_for_expensive(Node *n, Node* earliest) { 150 assert(n->in(0) && n->is_expensive(), "expensive node with control input here"); 151 assert(OptimizeExpensiveOps, "optimization off?"); 152 153 Node* ctl = n->in(0); 154 assert(ctl->is_CFG(), "expensive input 0 must be cfg"); 155 uint min_dom_depth = dom_depth(earliest); 156 #ifdef ASSERT 157 if (!is_dominator(ctl, earliest) && !is_dominator(earliest, ctl)) { 158 dump_bad_graph("Bad graph detected in get_early_ctrl_for_expensive", n, earliest, ctl); 159 assert(false, "Bad graph detected in get_early_ctrl_for_expensive"); 160 } 161 #endif 162 if (dom_depth(ctl) < min_dom_depth) { 163 return earliest; 164 } 165 166 while (1) { 167 Node *next = ctl; 168 // Moving the node out of a loop on the projection of a If 169 // confuses loop predication. So once we hit a Loop in a If branch 170 // that doesn't branch to an UNC, we stop. The code that process 171 // expensive nodes will notice the loop and skip over it to try to 172 // move the node further up. 173 if (ctl->is_CountedLoop() && ctl->in(1) != NULL && ctl->in(1)->in(0) != NULL && ctl->in(1)->in(0)->is_If()) { 174 if (!ctl->in(1)->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 175 break; 176 } 177 next = idom(ctl->in(1)->in(0)); 178 } else if (ctl->is_Proj()) { 179 // We only move it up along a projection if the projection is 180 // the single control projection for its parent: same code path, 181 // if it's a If with UNC or fallthrough of a call. 182 Node* parent_ctl = ctl->in(0); 183 if (parent_ctl == NULL) { 184 break; 185 } else if (parent_ctl->is_CountedLoopEnd() && parent_ctl->as_CountedLoopEnd()->loopnode() != NULL) { 186 next = parent_ctl->as_CountedLoopEnd()->loopnode()->init_control(); 187 } else if (parent_ctl->is_If()) { 188 if (!ctl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none)) { 189 break; 190 } 191 assert(idom(ctl) == parent_ctl, "strange"); 192 next = idom(parent_ctl); 193 } else if (ctl->is_CatchProj()) { 194 if (ctl->as_Proj()->_con != CatchProjNode::fall_through_index) { 195 break; 196 } 197 assert(parent_ctl->in(0)->in(0)->is_Call(), "strange graph"); 198 next = parent_ctl->in(0)->in(0)->in(0); 199 } else { 200 // Check if parent control has a single projection (this 201 // control is the only possible successor of the parent 202 // control). If so, we can try to move the node above the 203 // parent control. 204 int nb_ctl_proj = 0; 205 for (DUIterator_Fast imax, i = parent_ctl->fast_outs(imax); i < imax; i++) { 206 Node *p = parent_ctl->fast_out(i); 207 if (p->is_Proj() && p->is_CFG()) { 208 nb_ctl_proj++; 209 if (nb_ctl_proj > 1) { 210 break; 211 } 212 } 213 } 214 215 if (nb_ctl_proj > 1) { 216 break; 217 } 218 assert(parent_ctl->is_Start() || parent_ctl->is_MemBar() || parent_ctl->is_Call() || 219 BarrierSet::barrier_set()->barrier_set_c2()->is_gc_barrier_node(parent_ctl), "unexpected node"); 220 assert(idom(ctl) == parent_ctl, "strange"); 221 next = idom(parent_ctl); 222 } 223 } else { 224 next = idom(ctl); 225 } 226 if (next->is_Root() || next->is_Start() || dom_depth(next) < min_dom_depth) { 227 break; 228 } 229 ctl = next; 230 } 231 232 if (ctl != n->in(0)) { 233 _igvn.replace_input_of(n, 0, ctl); 234 _igvn.hash_insert(n); 235 } 236 237 return ctl; 238 } 239 240 241 //------------------------------set_early_ctrl--------------------------------- 242 // Set earliest legal control 243 void PhaseIdealLoop::set_early_ctrl( Node *n ) { 244 Node *early = get_early_ctrl(n); 245 246 // Record earliest legal location 247 set_ctrl(n, early); 248 } 249 250 //------------------------------set_subtree_ctrl------------------------------- 251 // set missing _ctrl entries on new nodes 252 void PhaseIdealLoop::set_subtree_ctrl( Node *n ) { 253 // Already set? Get out. 254 if( _nodes[n->_idx] ) return; 255 // Recursively set _nodes array to indicate where the Node goes 256 uint i; 257 for( i = 0; i < n->req(); ++i ) { 258 Node *m = n->in(i); 259 if( m && m != C->root() ) 260 set_subtree_ctrl( m ); 261 } 262 263 // Fixup self 264 set_early_ctrl( n ); 265 } 266 267 // Create a skeleton strip mined outer loop: a Loop head before the 268 // inner strip mined loop, a safepoint and an exit condition guarded 269 // by an opaque node after the inner strip mined loop with a backedge 270 // to the loop head. The inner strip mined loop is left as it is. Only 271 // once loop optimizations are over, do we adjust the inner loop exit 272 // condition to limit its number of iterations, set the outer loop 273 // exit condition and add Phis to the outer loop head. Some loop 274 // optimizations that operate on the inner strip mined loop need to be 275 // aware of the outer strip mined loop: loop unswitching needs to 276 // clone the outer loop as well as the inner, unrolling needs to only 277 // clone the inner loop etc. No optimizations need to change the outer 278 // strip mined loop as it is only a skeleton. 279 IdealLoopTree* PhaseIdealLoop::create_outer_strip_mined_loop(BoolNode *test, Node *cmp, Node *init_control, 280 IdealLoopTree* loop, float cl_prob, float le_fcnt, 281 Node*& entry_control, Node*& iffalse) { 282 Node* outer_test = _igvn.intcon(0); 283 set_ctrl(outer_test, C->root()); 284 Node *orig = iffalse; 285 iffalse = iffalse->clone(); 286 _igvn.register_new_node_with_optimizer(iffalse); 287 set_idom(iffalse, idom(orig), dom_depth(orig)); 288 289 IfNode *outer_le = new OuterStripMinedLoopEndNode(iffalse, outer_test, cl_prob, le_fcnt); 290 Node *outer_ift = new IfTrueNode (outer_le); 291 Node* outer_iff = orig; 292 _igvn.replace_input_of(outer_iff, 0, outer_le); 293 294 LoopNode *outer_l = new OuterStripMinedLoopNode(C, init_control, outer_ift); 295 entry_control = outer_l; 296 297 IdealLoopTree* outer_ilt = new IdealLoopTree(this, outer_l, outer_ift); 298 IdealLoopTree* parent = loop->_parent; 299 IdealLoopTree* sibling = parent->_child; 300 if (sibling == loop) { 301 parent->_child = outer_ilt; 302 } else { 303 while (sibling->_next != loop) { 304 sibling = sibling->_next; 305 } 306 sibling->_next = outer_ilt; 307 } 308 outer_ilt->_next = loop->_next; 309 outer_ilt->_parent = parent; 310 outer_ilt->_child = loop; 311 outer_ilt->_nest = loop->_nest; 312 loop->_parent = outer_ilt; 313 loop->_next = NULL; 314 loop->_nest++; 315 316 set_loop(iffalse, outer_ilt); 317 register_control(outer_le, outer_ilt, iffalse); 318 register_control(outer_ift, outer_ilt, outer_le); 319 set_idom(outer_iff, outer_le, dom_depth(outer_le)); 320 _igvn.register_new_node_with_optimizer(outer_l); 321 set_loop(outer_l, outer_ilt); 322 set_idom(outer_l, init_control, dom_depth(init_control)+1); 323 324 return outer_ilt; 325 } 326 327 //------------------------------is_counted_loop-------------------------------- 328 bool PhaseIdealLoop::is_counted_loop(Node* x, IdealLoopTree*& loop) { 329 PhaseGVN *gvn = &_igvn; 330 331 // Counted loop head must be a good RegionNode with only 3 not NULL 332 // control input edges: Self, Entry, LoopBack. 333 if (x->in(LoopNode::Self) == NULL || x->req() != 3 || loop->_irreducible) { 334 return false; 335 } 336 Node *init_control = x->in(LoopNode::EntryControl); 337 Node *back_control = x->in(LoopNode::LoopBackControl); 338 if (init_control == NULL || back_control == NULL) // Partially dead 339 return false; 340 // Must also check for TOP when looking for a dead loop 341 if (init_control->is_top() || back_control->is_top()) 342 return false; 343 344 // Allow funny placement of Safepoint 345 if (back_control->Opcode() == Op_SafePoint) { 346 if (LoopStripMiningIter != 0) { 347 // Leaving the safepoint on the backedge and creating a 348 // CountedLoop will confuse optimizations. We can't move the 349 // safepoint around because its jvm state wouldn't match a new 350 // location. Give up on that loop. 351 return false; 352 } 353 back_control = back_control->in(TypeFunc::Control); 354 } 355 356 // Controlling test for loop 357 Node *iftrue = back_control; 358 uint iftrue_op = iftrue->Opcode(); 359 if (iftrue_op != Op_IfTrue && 360 iftrue_op != Op_IfFalse) 361 // I have a weird back-control. Probably the loop-exit test is in 362 // the middle of the loop and I am looking at some trailing control-flow 363 // merge point. To fix this I would have to partially peel the loop. 364 return false; // Obscure back-control 365 366 // Get boolean guarding loop-back test 367 Node *iff = iftrue->in(0); 368 if (get_loop(iff) != loop || !iff->in(1)->is_Bool()) 369 return false; 370 BoolNode *test = iff->in(1)->as_Bool(); 371 BoolTest::mask bt = test->_test._test; 372 float cl_prob = iff->as_If()->_prob; 373 if (iftrue_op == Op_IfFalse) { 374 bt = BoolTest(bt).negate(); 375 cl_prob = 1.0 - cl_prob; 376 } 377 // Get backedge compare 378 Node *cmp = test->in(1); 379 int cmp_op = cmp->Opcode(); 380 if (cmp_op != Op_CmpI) 381 return false; // Avoid pointer & float compares 382 383 // Find the trip-counter increment & limit. Limit must be loop invariant. 384 Node *incr = cmp->in(1); 385 Node *limit = cmp->in(2); 386 387 // --------- 388 // need 'loop()' test to tell if limit is loop invariant 389 // --------- 390 391 if (!is_member(loop, get_ctrl(incr))) { // Swapped trip counter and limit? 392 Node *tmp = incr; // Then reverse order into the CmpI 393 incr = limit; 394 limit = tmp; 395 bt = BoolTest(bt).commute(); // And commute the exit test 396 } 397 if (is_member(loop, get_ctrl(limit))) // Limit must be loop-invariant 398 return false; 399 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 400 return false; 401 402 Node* phi_incr = NULL; 403 // Trip-counter increment must be commutative & associative. 404 if (incr->Opcode() == Op_CastII) { 405 incr = incr->in(1); 406 } 407 if (incr->is_Phi()) { 408 if (incr->as_Phi()->region() != x || incr->req() != 3) 409 return false; // Not simple trip counter expression 410 phi_incr = incr; 411 incr = phi_incr->in(LoopNode::LoopBackControl); // Assume incr is on backedge of Phi 412 if (!is_member(loop, get_ctrl(incr))) // Trip counter must be loop-variant 413 return false; 414 } 415 416 Node* trunc1 = NULL; 417 Node* trunc2 = NULL; 418 const TypeInt* iv_trunc_t = NULL; 419 Node* orig_incr = incr; 420 if (!(incr = CountedLoopNode::match_incr_with_optional_truncation(incr, &trunc1, &trunc2, &iv_trunc_t))) { 421 return false; // Funny increment opcode 422 } 423 assert(incr->Opcode() == Op_AddI, "wrong increment code"); 424 425 const TypeInt* limit_t = gvn->type(limit)->is_int(); 426 if (trunc1 != NULL) { 427 // When there is a truncation, we must be sure that after the truncation 428 // the trip counter will end up higher than the limit, otherwise we are looking 429 // at an endless loop. Can happen with range checks. 430 431 // Example: 432 // int i = 0; 433 // while (true) 434 // sum + = array[i]; 435 // i++; 436 // i = i && 0x7fff; 437 // } 438 // 439 // If the array is shorter than 0x8000 this exits through a AIOOB 440 // - Counted loop transformation is ok 441 // If the array is longer then this is an endless loop 442 // - No transformation can be done. 443 444 const TypeInt* incr_t = gvn->type(orig_incr)->is_int(); 445 if (limit_t->_hi > incr_t->_hi) { 446 // if the limit can have a higher value than the increment (before the phi) 447 return false; 448 } 449 } 450 451 // Get merge point 452 Node *xphi = incr->in(1); 453 Node *stride = incr->in(2); 454 if (!stride->is_Con()) { // Oops, swap these 455 if (!xphi->is_Con()) // Is the other guy a constant? 456 return false; // Nope, unknown stride, bail out 457 Node *tmp = xphi; // 'incr' is commutative, so ok to swap 458 xphi = stride; 459 stride = tmp; 460 } 461 if (xphi->Opcode() == Op_CastII) { 462 xphi = xphi->in(1); 463 } 464 // Stride must be constant 465 int stride_con = stride->get_int(); 466 if (stride_con == 0) 467 return false; // missed some peephole opt 468 469 if (!xphi->is_Phi()) 470 return false; // Too much math on the trip counter 471 if (phi_incr != NULL && phi_incr != xphi) 472 return false; 473 PhiNode *phi = xphi->as_Phi(); 474 475 // Phi must be of loop header; backedge must wrap to increment 476 if (phi->region() != x) 477 return false; 478 if ((trunc1 == NULL && phi->in(LoopNode::LoopBackControl) != incr) || 479 (trunc1 != NULL && phi->in(LoopNode::LoopBackControl) != trunc1)) { 480 return false; 481 } 482 Node *init_trip = phi->in(LoopNode::EntryControl); 483 484 // If iv trunc type is smaller than int, check for possible wrap. 485 if (!TypeInt::INT->higher_equal(iv_trunc_t)) { 486 assert(trunc1 != NULL, "must have found some truncation"); 487 488 // Get a better type for the phi (filtered thru if's) 489 const TypeInt* phi_ft = filtered_type(phi); 490 491 // Can iv take on a value that will wrap? 492 // 493 // Ensure iv's limit is not within "stride" of the wrap value. 494 // 495 // Example for "short" type 496 // Truncation ensures value is in the range -32768..32767 (iv_trunc_t) 497 // If the stride is +10, then the last value of the induction 498 // variable before the increment (phi_ft->_hi) must be 499 // <= 32767 - 10 and (phi_ft->_lo) must be >= -32768 to 500 // ensure no truncation occurs after the increment. 501 502 if (stride_con > 0) { 503 if (iv_trunc_t->_hi - phi_ft->_hi < stride_con || 504 iv_trunc_t->_lo > phi_ft->_lo) { 505 return false; // truncation may occur 506 } 507 } else if (stride_con < 0) { 508 if (iv_trunc_t->_lo - phi_ft->_lo > stride_con || 509 iv_trunc_t->_hi < phi_ft->_hi) { 510 return false; // truncation may occur 511 } 512 } 513 // No possibility of wrap so truncation can be discarded 514 // Promote iv type to Int 515 } else { 516 assert(trunc1 == NULL && trunc2 == NULL, "no truncation for int"); 517 } 518 519 // If the condition is inverted and we will be rolling 520 // through MININT to MAXINT, then bail out. 521 if (bt == BoolTest::eq || // Bail out, but this loop trips at most twice! 522 // Odd stride 523 (bt == BoolTest::ne && stride_con != 1 && stride_con != -1) || 524 // Count down loop rolls through MAXINT 525 ((bt == BoolTest::le || bt == BoolTest::lt) && stride_con < 0) || 526 // Count up loop rolls through MININT 527 ((bt == BoolTest::ge || bt == BoolTest::gt) && stride_con > 0)) { 528 return false; // Bail out 529 } 530 531 const TypeInt* init_t = gvn->type(init_trip)->is_int(); 532 533 if (stride_con > 0) { 534 jlong init_p = (jlong)init_t->_lo + stride_con; 535 if (init_p > (jlong)max_jint || init_p > (jlong)limit_t->_hi) 536 return false; // cyclic loop or this loop trips only once 537 } else { 538 jlong init_p = (jlong)init_t->_hi + stride_con; 539 if (init_p < (jlong)min_jint || init_p < (jlong)limit_t->_lo) 540 return false; // cyclic loop or this loop trips only once 541 } 542 543 if (phi_incr != NULL) { 544 // check if there is a possiblity of IV overflowing after the first increment 545 if (stride_con > 0) { 546 if (init_t->_hi > max_jint - stride_con) { 547 return false; 548 } 549 } else { 550 if (init_t->_lo < min_jint - stride_con) { 551 return false; 552 } 553 } 554 } 555 556 // ================================================= 557 // ---- SUCCESS! Found A Trip-Counted Loop! ----- 558 // 559 assert(x->Opcode() == Op_Loop, "regular loops only"); 560 C->print_method(PHASE_BEFORE_CLOOPS, 3); 561 562 Node *hook = new Node(6); 563 564 // =================================================== 565 // Generate loop limit check to avoid integer overflow 566 // in cases like next (cyclic loops): 567 // 568 // for (i=0; i <= max_jint; i++) {} 569 // for (i=0; i < max_jint; i+=2) {} 570 // 571 // 572 // Limit check predicate depends on the loop test: 573 // 574 // for(;i != limit; i++) --> limit <= (max_jint) 575 // for(;i < limit; i+=stride) --> limit <= (max_jint - stride + 1) 576 // for(;i <= limit; i+=stride) --> limit <= (max_jint - stride ) 577 // 578 579 // Check if limit is excluded to do more precise int overflow check. 580 bool incl_limit = (bt == BoolTest::le || bt == BoolTest::ge); 581 int stride_m = stride_con - (incl_limit ? 0 : (stride_con > 0 ? 1 : -1)); 582 583 // If compare points directly to the phi we need to adjust 584 // the compare so that it points to the incr. Limit have 585 // to be adjusted to keep trip count the same and the 586 // adjusted limit should be checked for int overflow. 587 if (phi_incr != NULL) { 588 stride_m += stride_con; 589 } 590 591 if (limit->is_Con()) { 592 int limit_con = limit->get_int(); 593 if ((stride_con > 0 && limit_con > (max_jint - stride_m)) || 594 (stride_con < 0 && limit_con < (min_jint - stride_m))) { 595 // Bailout: it could be integer overflow. 596 return false; 597 } 598 } else if ((stride_con > 0 && limit_t->_hi <= (max_jint - stride_m)) || 599 (stride_con < 0 && limit_t->_lo >= (min_jint - stride_m))) { 600 // Limit's type may satisfy the condition, for example, 601 // when it is an array length. 602 } else { 603 // Generate loop's limit check. 604 // Loop limit check predicate should be near the loop. 605 ProjNode *limit_check_proj = find_predicate_insertion_point(init_control, Deoptimization::Reason_loop_limit_check); 606 if (!limit_check_proj) { 607 // The limit check predicate is not generated if this method trapped here before. 608 #ifdef ASSERT 609 if (TraceLoopLimitCheck) { 610 tty->print("missing loop limit check:"); 611 loop->dump_head(); 612 x->dump(1); 613 } 614 #endif 615 return false; 616 } 617 618 IfNode* check_iff = limit_check_proj->in(0)->as_If(); 619 620 if (!is_dominator(get_ctrl(limit), check_iff->in(0))) { 621 return false; 622 } 623 624 Node* cmp_limit; 625 Node* bol; 626 627 if (stride_con > 0) { 628 cmp_limit = new CmpINode(limit, _igvn.intcon(max_jint - stride_m)); 629 bol = new BoolNode(cmp_limit, BoolTest::le); 630 } else { 631 cmp_limit = new CmpINode(limit, _igvn.intcon(min_jint - stride_m)); 632 bol = new BoolNode(cmp_limit, BoolTest::ge); 633 } 634 cmp_limit = _igvn.register_new_node_with_optimizer(cmp_limit); 635 bol = _igvn.register_new_node_with_optimizer(bol); 636 set_subtree_ctrl(bol); 637 638 // Replace condition in original predicate but preserve Opaque node 639 // so that previous predicates could be found. 640 assert(check_iff->in(1)->Opcode() == Op_Conv2B && 641 check_iff->in(1)->in(1)->Opcode() == Op_Opaque1, ""); 642 Node* opq = check_iff->in(1)->in(1); 643 _igvn.replace_input_of(opq, 1, bol); 644 // Update ctrl. 645 set_ctrl(opq, check_iff->in(0)); 646 set_ctrl(check_iff->in(1), check_iff->in(0)); 647 648 #ifndef PRODUCT 649 // report that the loop predication has been actually performed 650 // for this loop 651 if (TraceLoopLimitCheck) { 652 tty->print_cr("Counted Loop Limit Check generated:"); 653 debug_only( bol->dump(2); ) 654 } 655 #endif 656 } 657 658 if (phi_incr != NULL) { 659 // If compare points directly to the phi we need to adjust 660 // the compare so that it points to the incr. Limit have 661 // to be adjusted to keep trip count the same and we 662 // should avoid int overflow. 663 // 664 // i = init; do {} while(i++ < limit); 665 // is converted to 666 // i = init; do {} while(++i < limit+1); 667 // 668 limit = gvn->transform(new AddINode(limit, stride)); 669 } 670 671 // Now we need to canonicalize loop condition. 672 if (bt == BoolTest::ne) { 673 assert(stride_con == 1 || stride_con == -1, "simple increment only"); 674 // 'ne' can be replaced with 'lt' only when init < limit. 675 if (stride_con > 0 && init_t->_hi < limit_t->_lo) 676 bt = BoolTest::lt; 677 // 'ne' can be replaced with 'gt' only when init > limit. 678 if (stride_con < 0 && init_t->_lo > limit_t->_hi) 679 bt = BoolTest::gt; 680 } 681 682 if (incl_limit) { 683 // The limit check guaranties that 'limit <= (max_jint - stride)' so 684 // we can convert 'i <= limit' to 'i < limit+1' since stride != 0. 685 // 686 Node* one = (stride_con > 0) ? gvn->intcon( 1) : gvn->intcon(-1); 687 limit = gvn->transform(new AddINode(limit, one)); 688 if (bt == BoolTest::le) 689 bt = BoolTest::lt; 690 else if (bt == BoolTest::ge) 691 bt = BoolTest::gt; 692 else 693 ShouldNotReachHere(); 694 } 695 set_subtree_ctrl( limit ); 696 697 if (LoopStripMiningIter == 0) { 698 // Check for SafePoint on backedge and remove 699 Node *sfpt = x->in(LoopNode::LoopBackControl); 700 if (sfpt->Opcode() == Op_SafePoint && is_deleteable_safept(sfpt)) { 701 lazy_replace( sfpt, iftrue ); 702 if (loop->_safepts != NULL) { 703 loop->_safepts->yank(sfpt); 704 } 705 loop->_tail = iftrue; 706 } 707 } 708 709 // Build a canonical trip test. 710 // Clone code, as old values may be in use. 711 incr = incr->clone(); 712 incr->set_req(1,phi); 713 incr->set_req(2,stride); 714 incr = _igvn.register_new_node_with_optimizer(incr); 715 set_early_ctrl( incr ); 716 _igvn.rehash_node_delayed(phi); 717 phi->set_req_X( LoopNode::LoopBackControl, incr, &_igvn ); 718 719 // If phi type is more restrictive than Int, raise to 720 // Int to prevent (almost) infinite recursion in igvn 721 // which can only handle integer types for constants or minint..maxint. 722 if (!TypeInt::INT->higher_equal(phi->bottom_type())) { 723 Node* nphi = PhiNode::make(phi->in(0), phi->in(LoopNode::EntryControl), TypeInt::INT); 724 nphi->set_req(LoopNode::LoopBackControl, phi->in(LoopNode::LoopBackControl)); 725 nphi = _igvn.register_new_node_with_optimizer(nphi); 726 set_ctrl(nphi, get_ctrl(phi)); 727 _igvn.replace_node(phi, nphi); 728 phi = nphi->as_Phi(); 729 } 730 cmp = cmp->clone(); 731 cmp->set_req(1,incr); 732 cmp->set_req(2,limit); 733 cmp = _igvn.register_new_node_with_optimizer(cmp); 734 set_ctrl(cmp, iff->in(0)); 735 736 test = test->clone()->as_Bool(); 737 (*(BoolTest*)&test->_test)._test = bt; 738 test->set_req(1,cmp); 739 _igvn.register_new_node_with_optimizer(test); 740 set_ctrl(test, iff->in(0)); 741 742 // Replace the old IfNode with a new LoopEndNode 743 Node *lex = _igvn.register_new_node_with_optimizer(new CountedLoopEndNode( iff->in(0), test, cl_prob, iff->as_If()->_fcnt )); 744 IfNode *le = lex->as_If(); 745 uint dd = dom_depth(iff); 746 set_idom(le, le->in(0), dd); // Update dominance for loop exit 747 set_loop(le, loop); 748 749 // Get the loop-exit control 750 Node *iffalse = iff->as_If()->proj_out(!(iftrue_op == Op_IfTrue)); 751 752 // Need to swap loop-exit and loop-back control? 753 if (iftrue_op == Op_IfFalse) { 754 Node *ift2=_igvn.register_new_node_with_optimizer(new IfTrueNode (le)); 755 Node *iff2=_igvn.register_new_node_with_optimizer(new IfFalseNode(le)); 756 757 loop->_tail = back_control = ift2; 758 set_loop(ift2, loop); 759 set_loop(iff2, get_loop(iffalse)); 760 761 // Lazy update of 'get_ctrl' mechanism. 762 lazy_replace(iffalse, iff2); 763 lazy_replace(iftrue, ift2); 764 765 // Swap names 766 iffalse = iff2; 767 iftrue = ift2; 768 } else { 769 _igvn.rehash_node_delayed(iffalse); 770 _igvn.rehash_node_delayed(iftrue); 771 iffalse->set_req_X( 0, le, &_igvn ); 772 iftrue ->set_req_X( 0, le, &_igvn ); 773 } 774 775 set_idom(iftrue, le, dd+1); 776 set_idom(iffalse, le, dd+1); 777 assert(iff->outcnt() == 0, "should be dead now"); 778 lazy_replace( iff, le ); // fix 'get_ctrl' 779 780 Node *sfpt2 = le->in(0); 781 782 Node* entry_control = init_control; 783 bool strip_mine_loop = LoopStripMiningIter > 1 && loop->_child == NULL && 784 sfpt2->Opcode() == Op_SafePoint && !loop->_has_call; 785 IdealLoopTree* outer_ilt = NULL; 786 if (strip_mine_loop) { 787 outer_ilt = create_outer_strip_mined_loop(test, cmp, init_control, loop, 788 cl_prob, le->_fcnt, entry_control, 789 iffalse); 790 } 791 792 // Now setup a new CountedLoopNode to replace the existing LoopNode 793 CountedLoopNode *l = new CountedLoopNode(entry_control, back_control); 794 l->set_unswitch_count(x->as_Loop()->unswitch_count()); // Preserve 795 // The following assert is approximately true, and defines the intention 796 // of can_be_counted_loop. It fails, however, because phase->type 797 // is not yet initialized for this loop and its parts. 798 //assert(l->can_be_counted_loop(this), "sanity"); 799 _igvn.register_new_node_with_optimizer(l); 800 set_loop(l, loop); 801 loop->_head = l; 802 // Fix all data nodes placed at the old loop head. 803 // Uses the lazy-update mechanism of 'get_ctrl'. 804 lazy_replace( x, l ); 805 set_idom(l, entry_control, dom_depth(entry_control) + 1); 806 807 if (LoopStripMiningIter == 0 || strip_mine_loop) { 808 // Check for immediately preceding SafePoint and remove 809 if (sfpt2->Opcode() == Op_SafePoint && (LoopStripMiningIter != 0 || is_deleteable_safept(sfpt2))) { 810 if (strip_mine_loop) { 811 Node* outer_le = outer_ilt->_tail->in(0); 812 Node* sfpt = sfpt2->clone(); 813 sfpt->set_req(0, iffalse); 814 outer_le->set_req(0, sfpt); 815 register_control(sfpt, outer_ilt, iffalse); 816 set_idom(outer_le, sfpt, dom_depth(sfpt)); 817 } 818 lazy_replace( sfpt2, sfpt2->in(TypeFunc::Control)); 819 if (loop->_safepts != NULL) { 820 loop->_safepts->yank(sfpt2); 821 } 822 } 823 } 824 825 // Free up intermediate goo 826 _igvn.remove_dead_node(hook); 827 828 #ifdef ASSERT 829 assert(l->is_valid_counted_loop(), "counted loop shape is messed up"); 830 assert(l == loop->_head && l->phi() == phi && l->loopexit_or_null() == lex, "" ); 831 #endif 832 #ifndef PRODUCT 833 if (TraceLoopOpts) { 834 tty->print("Counted "); 835 loop->dump_head(); 836 } 837 #endif 838 839 C->print_method(PHASE_AFTER_CLOOPS, 3); 840 841 // Capture bounds of the loop in the induction variable Phi before 842 // subsequent transformation (iteration splitting) obscures the 843 // bounds 844 l->phi()->as_Phi()->set_type(l->phi()->Value(&_igvn)); 845 846 if (strip_mine_loop) { 847 l->mark_strip_mined(); 848 l->verify_strip_mined(1); 849 outer_ilt->_head->as_Loop()->verify_strip_mined(1); 850 loop = outer_ilt; 851 } 852 853 return true; 854 } 855 856 //----------------------exact_limit------------------------------------------- 857 Node* PhaseIdealLoop::exact_limit( IdealLoopTree *loop ) { 858 assert(loop->_head->is_CountedLoop(), ""); 859 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 860 assert(cl->is_valid_counted_loop(), ""); 861 862 if (ABS(cl->stride_con()) == 1 || 863 cl->limit()->Opcode() == Op_LoopLimit) { 864 // Old code has exact limit (it could be incorrect in case of int overflow). 865 // Loop limit is exact with stride == 1. And loop may already have exact limit. 866 return cl->limit(); 867 } 868 Node *limit = NULL; 869 #ifdef ASSERT 870 BoolTest::mask bt = cl->loopexit()->test_trip(); 871 assert(bt == BoolTest::lt || bt == BoolTest::gt, "canonical test is expected"); 872 #endif 873 if (cl->has_exact_trip_count()) { 874 // Simple case: loop has constant boundaries. 875 // Use jlongs to avoid integer overflow. 876 int stride_con = cl->stride_con(); 877 jlong init_con = cl->init_trip()->get_int(); 878 jlong limit_con = cl->limit()->get_int(); 879 julong trip_cnt = cl->trip_count(); 880 jlong final_con = init_con + trip_cnt*stride_con; 881 int final_int = (int)final_con; 882 // The final value should be in integer range since the loop 883 // is counted and the limit was checked for overflow. 884 assert(final_con == (jlong)final_int, "final value should be integer"); 885 limit = _igvn.intcon(final_int); 886 } else { 887 // Create new LoopLimit node to get exact limit (final iv value). 888 limit = new LoopLimitNode(C, cl->init_trip(), cl->limit(), cl->stride()); 889 register_new_node(limit, cl->in(LoopNode::EntryControl)); 890 } 891 assert(limit != NULL, "sanity"); 892 return limit; 893 } 894 895 //------------------------------Ideal------------------------------------------ 896 // Return a node which is more "ideal" than the current node. 897 // Attempt to convert into a counted-loop. 898 Node *LoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 899 if (!can_be_counted_loop(phase) && !is_OuterStripMinedLoop()) { 900 phase->C->set_major_progress(); 901 } 902 return RegionNode::Ideal(phase, can_reshape); 903 } 904 905 void LoopNode::verify_strip_mined(int expect_skeleton) const { 906 #ifdef ASSERT 907 const OuterStripMinedLoopNode* outer = NULL; 908 const CountedLoopNode* inner = NULL; 909 if (is_strip_mined()) { 910 assert(is_CountedLoop(), "no Loop should be marked strip mined"); 911 inner = as_CountedLoop(); 912 outer = inner->in(LoopNode::EntryControl)->as_OuterStripMinedLoop(); 913 } else if (is_OuterStripMinedLoop()) { 914 outer = this->as_OuterStripMinedLoop(); 915 inner = outer->unique_ctrl_out()->as_CountedLoop(); 916 assert(!is_strip_mined(), "outer loop shouldn't be marked strip mined"); 917 } 918 if (inner != NULL || outer != NULL) { 919 assert(inner != NULL && outer != NULL, "missing loop in strip mined nest"); 920 Node* outer_tail = outer->in(LoopNode::LoopBackControl); 921 Node* outer_le = outer_tail->in(0); 922 assert(outer_le->Opcode() == Op_OuterStripMinedLoopEnd, "tail of outer loop should be an If"); 923 Node* sfpt = outer_le->in(0); 924 assert(sfpt->Opcode() == Op_SafePoint, "where's the safepoint?"); 925 Node* inner_out = sfpt->in(0); 926 if (inner_out->outcnt() != 1) { 927 ResourceMark rm; 928 Unique_Node_List wq; 929 930 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 931 Node* u = inner_out->fast_out(i); 932 if (u == sfpt) { 933 continue; 934 } 935 wq.clear(); 936 wq.push(u); 937 bool found_sfpt = false; 938 for (uint next = 0; next < wq.size() && !found_sfpt; next++) { 939 Node *n = wq.at(next); 940 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && !found_sfpt; i++) { 941 Node* u = n->fast_out(i); 942 if (u == sfpt) { 943 found_sfpt = true; 944 } 945 if (!u->is_CFG()) { 946 wq.push(u); 947 } 948 } 949 } 950 assert(found_sfpt, "no node in loop that's not input to safepoint"); 951 } 952 } 953 CountedLoopEndNode* cle = inner_out->in(0)->as_CountedLoopEnd(); 954 assert(cle == inner->loopexit_or_null(), "mismatch"); 955 bool has_skeleton = outer_le->in(1)->bottom_type()->singleton() && outer_le->in(1)->bottom_type()->is_int()->get_con() == 0; 956 if (has_skeleton) { 957 assert(expect_skeleton == 1 || expect_skeleton == -1, "unexpected skeleton node"); 958 assert(outer->outcnt() == 2, "only phis"); 959 } else { 960 assert(expect_skeleton == 0 || expect_skeleton == -1, "no skeleton node?"); 961 uint phis = 0; 962 for (DUIterator_Fast imax, i = inner->fast_outs(imax); i < imax; i++) { 963 Node* u = inner->fast_out(i); 964 if (u->is_Phi()) { 965 phis++; 966 } 967 } 968 for (DUIterator_Fast imax, i = outer->fast_outs(imax); i < imax; i++) { 969 Node* u = outer->fast_out(i); 970 assert(u == outer || u == inner || u->is_Phi(), "nothing between inner and outer loop"); 971 } 972 uint stores = 0; 973 for (DUIterator_Fast imax, i = inner_out->fast_outs(imax); i < imax; i++) { 974 Node* u = inner_out->fast_out(i); 975 if (u->is_Store()) { 976 stores++; 977 } 978 } 979 assert(outer->outcnt() >= phis + 2 && outer->outcnt() <= phis + 2 + stores + 1, "only phis"); 980 } 981 assert(sfpt->outcnt() == 1, "no data node"); 982 assert(outer_tail->outcnt() == 1 || !has_skeleton, "no data node"); 983 } 984 #endif 985 } 986 987 //============================================================================= 988 //------------------------------Ideal------------------------------------------ 989 // Return a node which is more "ideal" than the current node. 990 // Attempt to convert into a counted-loop. 991 Node *CountedLoopNode::Ideal(PhaseGVN *phase, bool can_reshape) { 992 return RegionNode::Ideal(phase, can_reshape); 993 } 994 995 //------------------------------dump_spec-------------------------------------- 996 // Dump special per-node info 997 #ifndef PRODUCT 998 void CountedLoopNode::dump_spec(outputStream *st) const { 999 LoopNode::dump_spec(st); 1000 if (stride_is_con()) { 1001 st->print("stride: %d ",stride_con()); 1002 } 1003 if (is_pre_loop ()) st->print("pre of N%d" , _main_idx); 1004 if (is_main_loop()) st->print("main of N%d", _idx); 1005 if (is_post_loop()) st->print("post of N%d", _main_idx); 1006 if (is_strip_mined()) st->print(" strip mined"); 1007 } 1008 #endif 1009 1010 //============================================================================= 1011 int CountedLoopEndNode::stride_con() const { 1012 return stride()->bottom_type()->is_int()->get_con(); 1013 } 1014 1015 //============================================================================= 1016 //------------------------------Value----------------------------------------- 1017 const Type* LoopLimitNode::Value(PhaseGVN* phase) const { 1018 const Type* init_t = phase->type(in(Init)); 1019 const Type* limit_t = phase->type(in(Limit)); 1020 const Type* stride_t = phase->type(in(Stride)); 1021 // Either input is TOP ==> the result is TOP 1022 if (init_t == Type::TOP) return Type::TOP; 1023 if (limit_t == Type::TOP) return Type::TOP; 1024 if (stride_t == Type::TOP) return Type::TOP; 1025 1026 int stride_con = stride_t->is_int()->get_con(); 1027 if (stride_con == 1) 1028 return NULL; // Identity 1029 1030 if (init_t->is_int()->is_con() && limit_t->is_int()->is_con()) { 1031 // Use jlongs to avoid integer overflow. 1032 jlong init_con = init_t->is_int()->get_con(); 1033 jlong limit_con = limit_t->is_int()->get_con(); 1034 int stride_m = stride_con - (stride_con > 0 ? 1 : -1); 1035 jlong trip_count = (limit_con - init_con + stride_m)/stride_con; 1036 jlong final_con = init_con + stride_con*trip_count; 1037 int final_int = (int)final_con; 1038 // The final value should be in integer range since the loop 1039 // is counted and the limit was checked for overflow. 1040 assert(final_con == (jlong)final_int, "final value should be integer"); 1041 return TypeInt::make(final_int); 1042 } 1043 1044 return bottom_type(); // TypeInt::INT 1045 } 1046 1047 //------------------------------Ideal------------------------------------------ 1048 // Return a node which is more "ideal" than the current node. 1049 Node *LoopLimitNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1050 if (phase->type(in(Init)) == Type::TOP || 1051 phase->type(in(Limit)) == Type::TOP || 1052 phase->type(in(Stride)) == Type::TOP) 1053 return NULL; // Dead 1054 1055 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1056 if (stride_con == 1) 1057 return NULL; // Identity 1058 1059 if (in(Init)->is_Con() && in(Limit)->is_Con()) 1060 return NULL; // Value 1061 1062 // Delay following optimizations until all loop optimizations 1063 // done to keep Ideal graph simple. 1064 if (!can_reshape || phase->C->major_progress()) 1065 return NULL; 1066 1067 const TypeInt* init_t = phase->type(in(Init) )->is_int(); 1068 const TypeInt* limit_t = phase->type(in(Limit))->is_int(); 1069 int stride_p; 1070 jlong lim, ini; 1071 julong max; 1072 if (stride_con > 0) { 1073 stride_p = stride_con; 1074 lim = limit_t->_hi; 1075 ini = init_t->_lo; 1076 max = (julong)max_jint; 1077 } else { 1078 stride_p = -stride_con; 1079 lim = init_t->_hi; 1080 ini = limit_t->_lo; 1081 max = (julong)min_jint; 1082 } 1083 julong range = lim - ini + stride_p; 1084 if (range <= max) { 1085 // Convert to integer expression if it is not overflow. 1086 Node* stride_m = phase->intcon(stride_con - (stride_con > 0 ? 1 : -1)); 1087 Node *range = phase->transform(new SubINode(in(Limit), in(Init))); 1088 Node *bias = phase->transform(new AddINode(range, stride_m)); 1089 Node *trip = phase->transform(new DivINode(0, bias, in(Stride))); 1090 Node *span = phase->transform(new MulINode(trip, in(Stride))); 1091 return new AddINode(span, in(Init)); // exact limit 1092 } 1093 1094 if (is_power_of_2(stride_p) || // divisor is 2^n 1095 !Matcher::has_match_rule(Op_LoopLimit)) { // or no specialized Mach node? 1096 // Convert to long expression to avoid integer overflow 1097 // and let igvn optimizer convert this division. 1098 // 1099 Node* init = phase->transform( new ConvI2LNode(in(Init))); 1100 Node* limit = phase->transform( new ConvI2LNode(in(Limit))); 1101 Node* stride = phase->longcon(stride_con); 1102 Node* stride_m = phase->longcon(stride_con - (stride_con > 0 ? 1 : -1)); 1103 1104 Node *range = phase->transform(new SubLNode(limit, init)); 1105 Node *bias = phase->transform(new AddLNode(range, stride_m)); 1106 Node *span; 1107 if (stride_con > 0 && is_power_of_2(stride_p)) { 1108 // bias >= 0 if stride >0, so if stride is 2^n we can use &(-stride) 1109 // and avoid generating rounding for division. Zero trip guard should 1110 // guarantee that init < limit but sometimes the guard is missing and 1111 // we can get situation when init > limit. Note, for the empty loop 1112 // optimization zero trip guard is generated explicitly which leaves 1113 // only RCE predicate where exact limit is used and the predicate 1114 // will simply fail forcing recompilation. 1115 Node* neg_stride = phase->longcon(-stride_con); 1116 span = phase->transform(new AndLNode(bias, neg_stride)); 1117 } else { 1118 Node *trip = phase->transform(new DivLNode(0, bias, stride)); 1119 span = phase->transform(new MulLNode(trip, stride)); 1120 } 1121 // Convert back to int 1122 Node *span_int = phase->transform(new ConvL2INode(span)); 1123 return new AddINode(span_int, in(Init)); // exact limit 1124 } 1125 1126 return NULL; // No progress 1127 } 1128 1129 //------------------------------Identity--------------------------------------- 1130 // If stride == 1 return limit node. 1131 Node* LoopLimitNode::Identity(PhaseGVN* phase) { 1132 int stride_con = phase->type(in(Stride))->is_int()->get_con(); 1133 if (stride_con == 1 || stride_con == -1) 1134 return in(Limit); 1135 return this; 1136 } 1137 1138 //============================================================================= 1139 //----------------------match_incr_with_optional_truncation-------------------- 1140 // Match increment with optional truncation: 1141 // CHAR: (i+1)&0x7fff, BYTE: ((i+1)<<8)>>8, or SHORT: ((i+1)<<16)>>16 1142 // Return NULL for failure. Success returns the increment node. 1143 Node* CountedLoopNode::match_incr_with_optional_truncation( 1144 Node* expr, Node** trunc1, Node** trunc2, const TypeInt** trunc_type) { 1145 // Quick cutouts: 1146 if (expr == NULL || expr->req() != 3) return NULL; 1147 1148 Node *t1 = NULL; 1149 Node *t2 = NULL; 1150 const TypeInt* trunc_t = TypeInt::INT; 1151 Node* n1 = expr; 1152 int n1op = n1->Opcode(); 1153 1154 // Try to strip (n1 & M) or (n1 << N >> N) from n1. 1155 if (n1op == Op_AndI && 1156 n1->in(2)->is_Con() && 1157 n1->in(2)->bottom_type()->is_int()->get_con() == 0x7fff) { 1158 // %%% This check should match any mask of 2**K-1. 1159 t1 = n1; 1160 n1 = t1->in(1); 1161 n1op = n1->Opcode(); 1162 trunc_t = TypeInt::CHAR; 1163 } else if (n1op == Op_RShiftI && 1164 n1->in(1) != NULL && 1165 n1->in(1)->Opcode() == Op_LShiftI && 1166 n1->in(2) == n1->in(1)->in(2) && 1167 n1->in(2)->is_Con()) { 1168 jint shift = n1->in(2)->bottom_type()->is_int()->get_con(); 1169 // %%% This check should match any shift in [1..31]. 1170 if (shift == 16 || shift == 8) { 1171 t1 = n1; 1172 t2 = t1->in(1); 1173 n1 = t2->in(1); 1174 n1op = n1->Opcode(); 1175 if (shift == 16) { 1176 trunc_t = TypeInt::SHORT; 1177 } else if (shift == 8) { 1178 trunc_t = TypeInt::BYTE; 1179 } 1180 } 1181 } 1182 1183 // If (maybe after stripping) it is an AddI, we won: 1184 if (n1op == Op_AddI) { 1185 *trunc1 = t1; 1186 *trunc2 = t2; 1187 *trunc_type = trunc_t; 1188 return n1; 1189 } 1190 1191 // failed 1192 return NULL; 1193 } 1194 1195 LoopNode* CountedLoopNode::skip_strip_mined(int expect_skeleton) { 1196 if (is_strip_mined()) { 1197 verify_strip_mined(expect_skeleton); 1198 return in(EntryControl)->as_Loop(); 1199 } 1200 return this; 1201 } 1202 1203 OuterStripMinedLoopNode* CountedLoopNode::outer_loop() const { 1204 assert(is_strip_mined(), "not a strip mined loop"); 1205 Node* c = in(EntryControl); 1206 if (c == NULL || c->is_top() || !c->is_OuterStripMinedLoop()) { 1207 return NULL; 1208 } 1209 return c->as_OuterStripMinedLoop(); 1210 } 1211 1212 IfTrueNode* OuterStripMinedLoopNode::outer_loop_tail() const { 1213 Node* c = in(LoopBackControl); 1214 if (c == NULL || c->is_top()) { 1215 return NULL; 1216 } 1217 return c->as_IfTrue(); 1218 } 1219 1220 IfTrueNode* CountedLoopNode::outer_loop_tail() const { 1221 LoopNode* l = outer_loop(); 1222 if (l == NULL) { 1223 return NULL; 1224 } 1225 return l->outer_loop_tail(); 1226 } 1227 1228 OuterStripMinedLoopEndNode* OuterStripMinedLoopNode::outer_loop_end() const { 1229 IfTrueNode* proj = outer_loop_tail(); 1230 if (proj == NULL) { 1231 return NULL; 1232 } 1233 Node* c = proj->in(0); 1234 if (c == NULL || c->is_top() || c->outcnt() != 2) { 1235 return NULL; 1236 } 1237 return c->as_OuterStripMinedLoopEnd(); 1238 } 1239 1240 OuterStripMinedLoopEndNode* CountedLoopNode::outer_loop_end() const { 1241 LoopNode* l = outer_loop(); 1242 if (l == NULL) { 1243 return NULL; 1244 } 1245 return l->outer_loop_end(); 1246 } 1247 1248 IfFalseNode* OuterStripMinedLoopNode::outer_loop_exit() const { 1249 IfNode* le = outer_loop_end(); 1250 if (le == NULL) { 1251 return NULL; 1252 } 1253 Node* c = le->proj_out_or_null(false); 1254 if (c == NULL) { 1255 return NULL; 1256 } 1257 return c->as_IfFalse(); 1258 } 1259 1260 IfFalseNode* CountedLoopNode::outer_loop_exit() const { 1261 LoopNode* l = outer_loop(); 1262 if (l == NULL) { 1263 return NULL; 1264 } 1265 return l->outer_loop_exit(); 1266 } 1267 1268 SafePointNode* OuterStripMinedLoopNode::outer_safepoint() const { 1269 IfNode* le = outer_loop_end(); 1270 if (le == NULL) { 1271 return NULL; 1272 } 1273 Node* c = le->in(0); 1274 if (c == NULL || c->is_top()) { 1275 return NULL; 1276 } 1277 assert(c->Opcode() == Op_SafePoint, "broken outer loop"); 1278 return c->as_SafePoint(); 1279 } 1280 1281 SafePointNode* CountedLoopNode::outer_safepoint() const { 1282 LoopNode* l = outer_loop(); 1283 if (l == NULL) { 1284 return NULL; 1285 } 1286 return l->outer_safepoint(); 1287 } 1288 1289 Node* CountedLoopNode::skip_predicates_from_entry(Node* ctrl) { 1290 while (ctrl != NULL && ctrl->is_Proj() && ctrl->in(0)->is_If() && 1291 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->outcnt() == 1 && 1292 ctrl->in(0)->as_If()->proj_out(1-ctrl->as_Proj()->_con)->unique_out()->Opcode() == Op_Halt) { 1293 ctrl = ctrl->in(0)->in(0); 1294 } 1295 1296 return ctrl; 1297 } 1298 1299 Node* CountedLoopNode::skip_predicates() { 1300 if (is_main_loop()) { 1301 Node* ctrl = skip_strip_mined()->in(LoopNode::EntryControl); 1302 1303 return skip_predicates_from_entry(ctrl); 1304 } 1305 return in(LoopNode::EntryControl); 1306 } 1307 1308 void OuterStripMinedLoopNode::adjust_strip_mined_loop(PhaseIterGVN* igvn) { 1309 // Look for the outer & inner strip mined loop, reduce number of 1310 // iterations of the inner loop, set exit condition of outer loop, 1311 // construct required phi nodes for outer loop. 1312 CountedLoopNode* inner_cl = unique_ctrl_out()->as_CountedLoop(); 1313 assert(inner_cl->is_strip_mined(), "inner loop should be strip mined"); 1314 Node* inner_iv_phi = inner_cl->phi(); 1315 if (inner_iv_phi == NULL) { 1316 IfNode* outer_le = outer_loop_end(); 1317 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1318 igvn->replace_node(outer_le, iff); 1319 inner_cl->clear_strip_mined(); 1320 return; 1321 } 1322 CountedLoopEndNode* inner_cle = inner_cl->loopexit(); 1323 1324 int stride = inner_cl->stride_con(); 1325 jlong scaled_iters_long = ((jlong)LoopStripMiningIter) * ABS(stride); 1326 int scaled_iters = (int)scaled_iters_long; 1327 int short_scaled_iters = LoopStripMiningIterShortLoop* ABS(stride); 1328 const TypeInt* inner_iv_t = igvn->type(inner_iv_phi)->is_int(); 1329 jlong iter_estimate = (jlong)inner_iv_t->_hi - (jlong)inner_iv_t->_lo; 1330 assert(iter_estimate > 0, "broken"); 1331 if ((jlong)scaled_iters != scaled_iters_long || iter_estimate <= short_scaled_iters) { 1332 // Remove outer loop and safepoint (too few iterations) 1333 Node* outer_sfpt = outer_safepoint(); 1334 Node* outer_out = outer_loop_exit(); 1335 igvn->replace_node(outer_out, outer_sfpt->in(0)); 1336 igvn->replace_input_of(outer_sfpt, 0, igvn->C->top()); 1337 inner_cl->clear_strip_mined(); 1338 return; 1339 } 1340 if (iter_estimate <= scaled_iters_long) { 1341 // We would only go through one iteration of 1342 // the outer loop: drop the outer loop but 1343 // keep the safepoint so we don't run for 1344 // too long without a safepoint 1345 IfNode* outer_le = outer_loop_end(); 1346 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1347 igvn->replace_node(outer_le, iff); 1348 inner_cl->clear_strip_mined(); 1349 return; 1350 } 1351 1352 Node* cle_tail = inner_cle->proj_out(true); 1353 ResourceMark rm; 1354 Node_List old_new; 1355 if (cle_tail->outcnt() > 1) { 1356 // Look for nodes on backedge of inner loop and clone them 1357 Unique_Node_List backedge_nodes; 1358 for (DUIterator_Fast imax, i = cle_tail->fast_outs(imax); i < imax; i++) { 1359 Node* u = cle_tail->fast_out(i); 1360 if (u != inner_cl) { 1361 assert(!u->is_CFG(), "control flow on the backedge?"); 1362 backedge_nodes.push(u); 1363 } 1364 } 1365 uint last = igvn->C->unique(); 1366 for (uint next = 0; next < backedge_nodes.size(); next++) { 1367 Node* n = backedge_nodes.at(next); 1368 old_new.map(n->_idx, n->clone()); 1369 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 1370 Node* u = n->fast_out(i); 1371 assert(!u->is_CFG(), "broken"); 1372 if (u->_idx >= last) { 1373 continue; 1374 } 1375 if (!u->is_Phi()) { 1376 backedge_nodes.push(u); 1377 } else { 1378 assert(u->in(0) == inner_cl, "strange phi on the backedge"); 1379 } 1380 } 1381 } 1382 // Put the clones on the outer loop backedge 1383 Node* le_tail = outer_loop_tail(); 1384 for (uint next = 0; next < backedge_nodes.size(); next++) { 1385 Node *n = old_new[backedge_nodes.at(next)->_idx]; 1386 for (uint i = 1; i < n->req(); i++) { 1387 if (n->in(i) != NULL && old_new[n->in(i)->_idx] != NULL) { 1388 n->set_req(i, old_new[n->in(i)->_idx]); 1389 } 1390 } 1391 if (n->in(0) != NULL && n->in(0) == cle_tail) { 1392 n->set_req(0, le_tail); 1393 } 1394 igvn->register_new_node_with_optimizer(n); 1395 } 1396 } 1397 1398 Node* iv_phi = NULL; 1399 // Make a clone of each phi in the inner loop 1400 // for the outer loop 1401 for (uint i = 0; i < inner_cl->outcnt(); i++) { 1402 Node* u = inner_cl->raw_out(i); 1403 if (u->is_Phi()) { 1404 assert(u->in(0) == inner_cl, "inconsistent"); 1405 Node* phi = u->clone(); 1406 phi->set_req(0, this); 1407 Node* be = old_new[phi->in(LoopNode::LoopBackControl)->_idx]; 1408 if (be != NULL) { 1409 phi->set_req(LoopNode::LoopBackControl, be); 1410 } 1411 phi = igvn->transform(phi); 1412 igvn->replace_input_of(u, LoopNode::EntryControl, phi); 1413 if (u == inner_iv_phi) { 1414 iv_phi = phi; 1415 } 1416 } 1417 } 1418 Node* cle_out = inner_cle->proj_out(false); 1419 if (cle_out->outcnt() > 1) { 1420 // Look for chains of stores that were sunk 1421 // out of the inner loop and are in the outer loop 1422 for (DUIterator_Fast imax, i = cle_out->fast_outs(imax); i < imax; i++) { 1423 Node* u = cle_out->fast_out(i); 1424 if (u->is_Store()) { 1425 Node* first = u; 1426 for(;;) { 1427 Node* next = first->in(MemNode::Memory); 1428 if (!next->is_Store() || next->in(0) != cle_out) { 1429 break; 1430 } 1431 first = next; 1432 } 1433 Node* last = u; 1434 for(;;) { 1435 Node* next = NULL; 1436 for (DUIterator_Fast jmax, j = last->fast_outs(jmax); j < jmax; j++) { 1437 Node* uu = last->fast_out(j); 1438 if (uu->is_Store() && uu->in(0) == cle_out) { 1439 assert(next == NULL, "only one in the outer loop"); 1440 next = uu; 1441 } 1442 } 1443 if (next == NULL) { 1444 break; 1445 } 1446 last = next; 1447 } 1448 Node* phi = NULL; 1449 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1450 Node* uu = fast_out(j); 1451 if (uu->is_Phi()) { 1452 Node* be = uu->in(LoopNode::LoopBackControl); 1453 if (be->is_Store() && old_new[be->_idx] != NULL) { 1454 assert(false, "store on the backedge + sunk stores: unsupported"); 1455 // drop outer loop 1456 IfNode* outer_le = outer_loop_end(); 1457 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1458 igvn->replace_node(outer_le, iff); 1459 inner_cl->clear_strip_mined(); 1460 return; 1461 } 1462 if (be == last || be == first->in(MemNode::Memory)) { 1463 assert(phi == NULL, "only one phi"); 1464 phi = uu; 1465 } 1466 } 1467 } 1468 #ifdef ASSERT 1469 for (DUIterator_Fast jmax, j = fast_outs(jmax); j < jmax; j++) { 1470 Node* uu = fast_out(j); 1471 if (uu->is_Phi() && uu->bottom_type() == Type::MEMORY) { 1472 if (uu->adr_type() == igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))) { 1473 assert(phi == uu, "what's that phi?"); 1474 } else if (uu->adr_type() == TypePtr::BOTTOM) { 1475 Node* n = uu->in(LoopNode::LoopBackControl); 1476 uint limit = igvn->C->live_nodes(); 1477 uint i = 0; 1478 while (n != uu) { 1479 i++; 1480 assert(i < limit, "infinite loop"); 1481 if (n->is_Proj()) { 1482 n = n->in(0); 1483 } else if (n->is_SafePoint() || n->is_MemBar()) { 1484 n = n->in(TypeFunc::Memory); 1485 } else if (n->is_Phi()) { 1486 n = n->in(1); 1487 } else if (n->is_MergeMem()) { 1488 n = n->as_MergeMem()->memory_at(igvn->C->get_alias_index(u->adr_type())); 1489 } else if (n->is_Store() || n->is_LoadStore() || n->is_ClearArray()) { 1490 n = n->in(MemNode::Memory); 1491 } else { 1492 n->dump(); 1493 ShouldNotReachHere(); 1494 } 1495 } 1496 } 1497 } 1498 } 1499 #endif 1500 if (phi == NULL) { 1501 // If the an entire chains was sunk, the 1502 // inner loop has no phi for that memory 1503 // slice, create one for the outer loop 1504 phi = PhiNode::make(this, first->in(MemNode::Memory), Type::MEMORY, 1505 igvn->C->get_adr_type(igvn->C->get_alias_index(u->adr_type()))); 1506 phi->set_req(LoopNode::LoopBackControl, last); 1507 phi = igvn->transform(phi); 1508 igvn->replace_input_of(first, MemNode::Memory, phi); 1509 } else { 1510 // Or fix the outer loop fix to include 1511 // that chain of stores. 1512 Node* be = phi->in(LoopNode::LoopBackControl); 1513 assert(!(be->is_Store() && old_new[be->_idx] != NULL), "store on the backedge + sunk stores: unsupported"); 1514 if (be == first->in(MemNode::Memory)) { 1515 if (be == phi->in(LoopNode::LoopBackControl)) { 1516 igvn->replace_input_of(phi, LoopNode::LoopBackControl, last); 1517 } else { 1518 igvn->replace_input_of(be, MemNode::Memory, last); 1519 } 1520 } else { 1521 #ifdef ASSERT 1522 if (be == phi->in(LoopNode::LoopBackControl)) { 1523 assert(phi->in(LoopNode::LoopBackControl) == last, ""); 1524 } else { 1525 assert(be->in(MemNode::Memory) == last, ""); 1526 } 1527 #endif 1528 } 1529 } 1530 } 1531 } 1532 } 1533 1534 if (iv_phi != NULL) { 1535 // Now adjust the inner loop's exit condition 1536 Node* limit = inner_cl->limit(); 1537 Node* sub = NULL; 1538 if (stride > 0) { 1539 sub = igvn->transform(new SubINode(limit, iv_phi)); 1540 } else { 1541 sub = igvn->transform(new SubINode(iv_phi, limit)); 1542 } 1543 Node* min = igvn->transform(new MinINode(sub, igvn->intcon(scaled_iters))); 1544 Node* new_limit = NULL; 1545 if (stride > 0) { 1546 new_limit = igvn->transform(new AddINode(min, iv_phi)); 1547 } else { 1548 new_limit = igvn->transform(new SubINode(iv_phi, min)); 1549 } 1550 Node* cmp = inner_cle->cmp_node()->clone(); 1551 igvn->replace_input_of(cmp, 2, new_limit); 1552 Node* bol = inner_cle->in(CountedLoopEndNode::TestValue)->clone(); 1553 cmp->set_req(2, limit); 1554 bol->set_req(1, igvn->transform(cmp)); 1555 igvn->replace_input_of(outer_loop_end(), 1, igvn->transform(bol)); 1556 } else { 1557 assert(false, "should be able to adjust outer loop"); 1558 IfNode* outer_le = outer_loop_end(); 1559 Node* iff = igvn->transform(new IfNode(outer_le->in(0), outer_le->in(1), outer_le->_prob, outer_le->_fcnt)); 1560 igvn->replace_node(outer_le, iff); 1561 inner_cl->clear_strip_mined(); 1562 } 1563 } 1564 1565 const Type* OuterStripMinedLoopEndNode::Value(PhaseGVN* phase) const { 1566 if (!in(0)) return Type::TOP; 1567 if (phase->type(in(0)) == Type::TOP) 1568 return Type::TOP; 1569 1570 return TypeTuple::IFBOTH; 1571 } 1572 1573 Node *OuterStripMinedLoopEndNode::Ideal(PhaseGVN *phase, bool can_reshape) { 1574 if (remove_dead_region(phase, can_reshape)) return this; 1575 1576 return NULL; 1577 } 1578 1579 //------------------------------filtered_type-------------------------------- 1580 // Return a type based on condition control flow 1581 // A successful return will be a type that is restricted due 1582 // to a series of dominating if-tests, such as: 1583 // if (i < 10) { 1584 // if (i > 0) { 1585 // here: "i" type is [1..10) 1586 // } 1587 // } 1588 // or a control flow merge 1589 // if (i < 10) { 1590 // do { 1591 // phi( , ) -- at top of loop type is [min_int..10) 1592 // i = ? 1593 // } while ( i < 10) 1594 // 1595 const TypeInt* PhaseIdealLoop::filtered_type( Node *n, Node* n_ctrl) { 1596 assert(n && n->bottom_type()->is_int(), "must be int"); 1597 const TypeInt* filtered_t = NULL; 1598 if (!n->is_Phi()) { 1599 assert(n_ctrl != NULL || n_ctrl == C->top(), "valid control"); 1600 filtered_t = filtered_type_from_dominators(n, n_ctrl); 1601 1602 } else { 1603 Node* phi = n->as_Phi(); 1604 Node* region = phi->in(0); 1605 assert(n_ctrl == NULL || n_ctrl == region, "ctrl parameter must be region"); 1606 if (region && region != C->top()) { 1607 for (uint i = 1; i < phi->req(); i++) { 1608 Node* val = phi->in(i); 1609 Node* use_c = region->in(i); 1610 const TypeInt* val_t = filtered_type_from_dominators(val, use_c); 1611 if (val_t != NULL) { 1612 if (filtered_t == NULL) { 1613 filtered_t = val_t; 1614 } else { 1615 filtered_t = filtered_t->meet(val_t)->is_int(); 1616 } 1617 } 1618 } 1619 } 1620 } 1621 const TypeInt* n_t = _igvn.type(n)->is_int(); 1622 if (filtered_t != NULL) { 1623 n_t = n_t->join(filtered_t)->is_int(); 1624 } 1625 return n_t; 1626 } 1627 1628 1629 //------------------------------filtered_type_from_dominators-------------------------------- 1630 // Return a possibly more restrictive type for val based on condition control flow of dominators 1631 const TypeInt* PhaseIdealLoop::filtered_type_from_dominators( Node* val, Node *use_ctrl) { 1632 if (val->is_Con()) { 1633 return val->bottom_type()->is_int(); 1634 } 1635 uint if_limit = 10; // Max number of dominating if's visited 1636 const TypeInt* rtn_t = NULL; 1637 1638 if (use_ctrl && use_ctrl != C->top()) { 1639 Node* val_ctrl = get_ctrl(val); 1640 uint val_dom_depth = dom_depth(val_ctrl); 1641 Node* pred = use_ctrl; 1642 uint if_cnt = 0; 1643 while (if_cnt < if_limit) { 1644 if ((pred->Opcode() == Op_IfTrue || pred->Opcode() == Op_IfFalse)) { 1645 if_cnt++; 1646 const TypeInt* if_t = IfNode::filtered_int_type(&_igvn, val, pred); 1647 if (if_t != NULL) { 1648 if (rtn_t == NULL) { 1649 rtn_t = if_t; 1650 } else { 1651 rtn_t = rtn_t->join(if_t)->is_int(); 1652 } 1653 } 1654 } 1655 pred = idom(pred); 1656 if (pred == NULL || pred == C->top()) { 1657 break; 1658 } 1659 // Stop if going beyond definition block of val 1660 if (dom_depth(pred) < val_dom_depth) { 1661 break; 1662 } 1663 } 1664 } 1665 return rtn_t; 1666 } 1667 1668 1669 //------------------------------dump_spec-------------------------------------- 1670 // Dump special per-node info 1671 #ifndef PRODUCT 1672 void CountedLoopEndNode::dump_spec(outputStream *st) const { 1673 if( in(TestValue) != NULL && in(TestValue)->is_Bool() ) { 1674 BoolTest bt( test_trip()); // Added this for g++. 1675 1676 st->print("["); 1677 bt.dump_on(st); 1678 st->print("]"); 1679 } 1680 st->print(" "); 1681 IfNode::dump_spec(st); 1682 } 1683 #endif 1684 1685 //============================================================================= 1686 //------------------------------is_member-------------------------------------- 1687 // Is 'l' a member of 'this'? 1688 bool IdealLoopTree::is_member(const IdealLoopTree *l) const { 1689 while( l->_nest > _nest ) l = l->_parent; 1690 return l == this; 1691 } 1692 1693 //------------------------------set_nest--------------------------------------- 1694 // Set loop tree nesting depth. Accumulate _has_call bits. 1695 int IdealLoopTree::set_nest( uint depth ) { 1696 _nest = depth; 1697 int bits = _has_call; 1698 if( _child ) bits |= _child->set_nest(depth+1); 1699 if( bits ) _has_call = 1; 1700 if( _next ) bits |= _next ->set_nest(depth ); 1701 return bits; 1702 } 1703 1704 //------------------------------split_fall_in---------------------------------- 1705 // Split out multiple fall-in edges from the loop header. Move them to a 1706 // private RegionNode before the loop. This becomes the loop landing pad. 1707 void IdealLoopTree::split_fall_in( PhaseIdealLoop *phase, int fall_in_cnt ) { 1708 PhaseIterGVN &igvn = phase->_igvn; 1709 uint i; 1710 1711 // Make a new RegionNode to be the landing pad. 1712 Node *landing_pad = new RegionNode( fall_in_cnt+1 ); 1713 phase->set_loop(landing_pad,_parent); 1714 // Gather all the fall-in control paths into the landing pad 1715 uint icnt = fall_in_cnt; 1716 uint oreq = _head->req(); 1717 for( i = oreq-1; i>0; i-- ) 1718 if( !phase->is_member( this, _head->in(i) ) ) 1719 landing_pad->set_req(icnt--,_head->in(i)); 1720 1721 // Peel off PhiNode edges as well 1722 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1723 Node *oj = _head->fast_out(j); 1724 if( oj->is_Phi() ) { 1725 PhiNode* old_phi = oj->as_Phi(); 1726 assert( old_phi->region() == _head, "" ); 1727 igvn.hash_delete(old_phi); // Yank from hash before hacking edges 1728 Node *p = PhiNode::make_blank(landing_pad, old_phi); 1729 uint icnt = fall_in_cnt; 1730 for( i = oreq-1; i>0; i-- ) { 1731 if( !phase->is_member( this, _head->in(i) ) ) { 1732 p->init_req(icnt--, old_phi->in(i)); 1733 // Go ahead and clean out old edges from old phi 1734 old_phi->del_req(i); 1735 } 1736 } 1737 // Search for CSE's here, because ZKM.jar does a lot of 1738 // loop hackery and we need to be a little incremental 1739 // with the CSE to avoid O(N^2) node blow-up. 1740 Node *p2 = igvn.hash_find_insert(p); // Look for a CSE 1741 if( p2 ) { // Found CSE 1742 p->destruct(); // Recover useless new node 1743 p = p2; // Use old node 1744 } else { 1745 igvn.register_new_node_with_optimizer(p, old_phi); 1746 } 1747 // Make old Phi refer to new Phi. 1748 old_phi->add_req(p); 1749 // Check for the special case of making the old phi useless and 1750 // disappear it. In JavaGrande I have a case where this useless 1751 // Phi is the loop limit and prevents recognizing a CountedLoop 1752 // which in turn prevents removing an empty loop. 1753 Node *id_old_phi = old_phi->Identity( &igvn ); 1754 if( id_old_phi != old_phi ) { // Found a simple identity? 1755 // Note that I cannot call 'replace_node' here, because 1756 // that will yank the edge from old_phi to the Region and 1757 // I'm mid-iteration over the Region's uses. 1758 for (DUIterator_Last imin, i = old_phi->last_outs(imin); i >= imin; ) { 1759 Node* use = old_phi->last_out(i); 1760 igvn.rehash_node_delayed(use); 1761 uint uses_found = 0; 1762 for (uint j = 0; j < use->len(); j++) { 1763 if (use->in(j) == old_phi) { 1764 if (j < use->req()) use->set_req (j, id_old_phi); 1765 else use->set_prec(j, id_old_phi); 1766 uses_found++; 1767 } 1768 } 1769 i -= uses_found; // we deleted 1 or more copies of this edge 1770 } 1771 } 1772 igvn._worklist.push(old_phi); 1773 } 1774 } 1775 // Finally clean out the fall-in edges from the RegionNode 1776 for( i = oreq-1; i>0; i-- ) { 1777 if( !phase->is_member( this, _head->in(i) ) ) { 1778 _head->del_req(i); 1779 } 1780 } 1781 igvn.rehash_node_delayed(_head); 1782 // Transform landing pad 1783 igvn.register_new_node_with_optimizer(landing_pad, _head); 1784 // Insert landing pad into the header 1785 _head->add_req(landing_pad); 1786 } 1787 1788 //------------------------------split_outer_loop------------------------------- 1789 // Split out the outermost loop from this shared header. 1790 void IdealLoopTree::split_outer_loop( PhaseIdealLoop *phase ) { 1791 PhaseIterGVN &igvn = phase->_igvn; 1792 1793 // Find index of outermost loop; it should also be my tail. 1794 uint outer_idx = 1; 1795 while( _head->in(outer_idx) != _tail ) outer_idx++; 1796 1797 // Make a LoopNode for the outermost loop. 1798 Node *ctl = _head->in(LoopNode::EntryControl); 1799 Node *outer = new LoopNode( ctl, _head->in(outer_idx) ); 1800 outer = igvn.register_new_node_with_optimizer(outer, _head); 1801 phase->set_created_loop_node(); 1802 1803 // Outermost loop falls into '_head' loop 1804 _head->set_req(LoopNode::EntryControl, outer); 1805 _head->del_req(outer_idx); 1806 // Split all the Phis up between '_head' loop and 'outer' loop. 1807 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1808 Node *out = _head->fast_out(j); 1809 if( out->is_Phi() ) { 1810 PhiNode *old_phi = out->as_Phi(); 1811 assert( old_phi->region() == _head, "" ); 1812 Node *phi = PhiNode::make_blank(outer, old_phi); 1813 phi->init_req(LoopNode::EntryControl, old_phi->in(LoopNode::EntryControl)); 1814 phi->init_req(LoopNode::LoopBackControl, old_phi->in(outer_idx)); 1815 phi = igvn.register_new_node_with_optimizer(phi, old_phi); 1816 // Make old Phi point to new Phi on the fall-in path 1817 igvn.replace_input_of(old_phi, LoopNode::EntryControl, phi); 1818 old_phi->del_req(outer_idx); 1819 } 1820 } 1821 1822 // Use the new loop head instead of the old shared one 1823 _head = outer; 1824 phase->set_loop(_head, this); 1825 } 1826 1827 //------------------------------fix_parent------------------------------------- 1828 static void fix_parent( IdealLoopTree *loop, IdealLoopTree *parent ) { 1829 loop->_parent = parent; 1830 if( loop->_child ) fix_parent( loop->_child, loop ); 1831 if( loop->_next ) fix_parent( loop->_next , parent ); 1832 } 1833 1834 //------------------------------estimate_path_freq----------------------------- 1835 static float estimate_path_freq( Node *n ) { 1836 // Try to extract some path frequency info 1837 IfNode *iff; 1838 for( int i = 0; i < 50; i++ ) { // Skip through a bunch of uncommon tests 1839 uint nop = n->Opcode(); 1840 if( nop == Op_SafePoint ) { // Skip any safepoint 1841 n = n->in(0); 1842 continue; 1843 } 1844 if( nop == Op_CatchProj ) { // Get count from a prior call 1845 // Assume call does not always throw exceptions: means the call-site 1846 // count is also the frequency of the fall-through path. 1847 assert( n->is_CatchProj(), "" ); 1848 if( ((CatchProjNode*)n)->_con != CatchProjNode::fall_through_index ) 1849 return 0.0f; // Assume call exception path is rare 1850 Node *call = n->in(0)->in(0)->in(0); 1851 assert( call->is_Call(), "expect a call here" ); 1852 const JVMState *jvms = ((CallNode*)call)->jvms(); 1853 ciMethodData* methodData = jvms->method()->method_data(); 1854 if (!methodData->is_mature()) return 0.0f; // No call-site data 1855 ciProfileData* data = methodData->bci_to_data(jvms->bci()); 1856 if ((data == NULL) || !data->is_CounterData()) { 1857 // no call profile available, try call's control input 1858 n = n->in(0); 1859 continue; 1860 } 1861 return data->as_CounterData()->count()/FreqCountInvocations; 1862 } 1863 // See if there's a gating IF test 1864 Node *n_c = n->in(0); 1865 if( !n_c->is_If() ) break; // No estimate available 1866 iff = n_c->as_If(); 1867 if( iff->_fcnt != COUNT_UNKNOWN ) // Have a valid count? 1868 // Compute how much count comes on this path 1869 return ((nop == Op_IfTrue) ? iff->_prob : 1.0f - iff->_prob) * iff->_fcnt; 1870 // Have no count info. Skip dull uncommon-trap like branches. 1871 if( (nop == Op_IfTrue && iff->_prob < PROB_LIKELY_MAG(5)) || 1872 (nop == Op_IfFalse && iff->_prob > PROB_UNLIKELY_MAG(5)) ) 1873 break; 1874 // Skip through never-taken branch; look for a real loop exit. 1875 n = iff->in(0); 1876 } 1877 return 0.0f; // No estimate available 1878 } 1879 1880 //------------------------------merge_many_backedges--------------------------- 1881 // Merge all the backedges from the shared header into a private Region. 1882 // Feed that region as the one backedge to this loop. 1883 void IdealLoopTree::merge_many_backedges( PhaseIdealLoop *phase ) { 1884 uint i; 1885 1886 // Scan for the top 2 hottest backedges 1887 float hotcnt = 0.0f; 1888 float warmcnt = 0.0f; 1889 uint hot_idx = 0; 1890 // Loop starts at 2 because slot 1 is the fall-in path 1891 for( i = 2; i < _head->req(); i++ ) { 1892 float cnt = estimate_path_freq(_head->in(i)); 1893 if( cnt > hotcnt ) { // Grab hottest path 1894 warmcnt = hotcnt; 1895 hotcnt = cnt; 1896 hot_idx = i; 1897 } else if( cnt > warmcnt ) { // And 2nd hottest path 1898 warmcnt = cnt; 1899 } 1900 } 1901 1902 // See if the hottest backedge is worthy of being an inner loop 1903 // by being much hotter than the next hottest backedge. 1904 if( hotcnt <= 0.0001 || 1905 hotcnt < 2.0*warmcnt ) hot_idx = 0;// No hot backedge 1906 1907 // Peel out the backedges into a private merge point; peel 1908 // them all except optionally hot_idx. 1909 PhaseIterGVN &igvn = phase->_igvn; 1910 1911 Node *hot_tail = NULL; 1912 // Make a Region for the merge point 1913 Node *r = new RegionNode(1); 1914 for( i = 2; i < _head->req(); i++ ) { 1915 if( i != hot_idx ) 1916 r->add_req( _head->in(i) ); 1917 else hot_tail = _head->in(i); 1918 } 1919 igvn.register_new_node_with_optimizer(r, _head); 1920 // Plug region into end of loop _head, followed by hot_tail 1921 while( _head->req() > 3 ) _head->del_req( _head->req()-1 ); 1922 igvn.replace_input_of(_head, 2, r); 1923 if( hot_idx ) _head->add_req(hot_tail); 1924 1925 // Split all the Phis up between '_head' loop and the Region 'r' 1926 for (DUIterator_Fast jmax, j = _head->fast_outs(jmax); j < jmax; j++) { 1927 Node *out = _head->fast_out(j); 1928 if( out->is_Phi() ) { 1929 PhiNode* n = out->as_Phi(); 1930 igvn.hash_delete(n); // Delete from hash before hacking edges 1931 Node *hot_phi = NULL; 1932 Node *phi = new PhiNode(r, n->type(), n->adr_type()); 1933 // Check all inputs for the ones to peel out 1934 uint j = 1; 1935 for( uint i = 2; i < n->req(); i++ ) { 1936 if( i != hot_idx ) 1937 phi->set_req( j++, n->in(i) ); 1938 else hot_phi = n->in(i); 1939 } 1940 // Register the phi but do not transform until whole place transforms 1941 igvn.register_new_node_with_optimizer(phi, n); 1942 // Add the merge phi to the old Phi 1943 while( n->req() > 3 ) n->del_req( n->req()-1 ); 1944 igvn.replace_input_of(n, 2, phi); 1945 if( hot_idx ) n->add_req(hot_phi); 1946 } 1947 } 1948 1949 1950 // Insert a new IdealLoopTree inserted below me. Turn it into a clone 1951 // of self loop tree. Turn self into a loop headed by _head and with 1952 // tail being the new merge point. 1953 IdealLoopTree *ilt = new IdealLoopTree( phase, _head, _tail ); 1954 phase->set_loop(_tail,ilt); // Adjust tail 1955 _tail = r; // Self's tail is new merge point 1956 phase->set_loop(r,this); 1957 ilt->_child = _child; // New guy has my children 1958 _child = ilt; // Self has new guy as only child 1959 ilt->_parent = this; // new guy has self for parent 1960 ilt->_nest = _nest; // Same nesting depth (for now) 1961 1962 // Starting with 'ilt', look for child loop trees using the same shared 1963 // header. Flatten these out; they will no longer be loops in the end. 1964 IdealLoopTree **pilt = &_child; 1965 while( ilt ) { 1966 if( ilt->_head == _head ) { 1967 uint i; 1968 for( i = 2; i < _head->req(); i++ ) 1969 if( _head->in(i) == ilt->_tail ) 1970 break; // Still a loop 1971 if( i == _head->req() ) { // No longer a loop 1972 // Flatten ilt. Hang ilt's "_next" list from the end of 1973 // ilt's '_child' list. Move the ilt's _child up to replace ilt. 1974 IdealLoopTree **cp = &ilt->_child; 1975 while( *cp ) cp = &(*cp)->_next; // Find end of child list 1976 *cp = ilt->_next; // Hang next list at end of child list 1977 *pilt = ilt->_child; // Move child up to replace ilt 1978 ilt->_head = NULL; // Flag as a loop UNIONED into parent 1979 ilt = ilt->_child; // Repeat using new ilt 1980 continue; // do not advance over ilt->_child 1981 } 1982 assert( ilt->_tail == hot_tail, "expected to only find the hot inner loop here" ); 1983 phase->set_loop(_head,ilt); 1984 } 1985 pilt = &ilt->_child; // Advance to next 1986 ilt = *pilt; 1987 } 1988 1989 if( _child ) fix_parent( _child, this ); 1990 } 1991 1992 //------------------------------beautify_loops--------------------------------- 1993 // Split shared headers and insert loop landing pads. 1994 // Insert a LoopNode to replace the RegionNode. 1995 // Return TRUE if loop tree is structurally changed. 1996 bool IdealLoopTree::beautify_loops( PhaseIdealLoop *phase ) { 1997 bool result = false; 1998 // Cache parts in locals for easy 1999 PhaseIterGVN &igvn = phase->_igvn; 2000 2001 igvn.hash_delete(_head); // Yank from hash before hacking edges 2002 2003 // Check for multiple fall-in paths. Peel off a landing pad if need be. 2004 int fall_in_cnt = 0; 2005 for( uint i = 1; i < _head->req(); i++ ) 2006 if( !phase->is_member( this, _head->in(i) ) ) 2007 fall_in_cnt++; 2008 assert( fall_in_cnt, "at least 1 fall-in path" ); 2009 if( fall_in_cnt > 1 ) // Need a loop landing pad to merge fall-ins 2010 split_fall_in( phase, fall_in_cnt ); 2011 2012 // Swap inputs to the _head and all Phis to move the fall-in edge to 2013 // the left. 2014 fall_in_cnt = 1; 2015 while( phase->is_member( this, _head->in(fall_in_cnt) ) ) 2016 fall_in_cnt++; 2017 if( fall_in_cnt > 1 ) { 2018 // Since I am just swapping inputs I do not need to update def-use info 2019 Node *tmp = _head->in(1); 2020 igvn.rehash_node_delayed(_head); 2021 _head->set_req( 1, _head->in(fall_in_cnt) ); 2022 _head->set_req( fall_in_cnt, tmp ); 2023 // Swap also all Phis 2024 for (DUIterator_Fast imax, i = _head->fast_outs(imax); i < imax; i++) { 2025 Node* phi = _head->fast_out(i); 2026 if( phi->is_Phi() ) { 2027 igvn.rehash_node_delayed(phi); // Yank from hash before hacking edges 2028 tmp = phi->in(1); 2029 phi->set_req( 1, phi->in(fall_in_cnt) ); 2030 phi->set_req( fall_in_cnt, tmp ); 2031 } 2032 } 2033 } 2034 assert( !phase->is_member( this, _head->in(1) ), "left edge is fall-in" ); 2035 assert( phase->is_member( this, _head->in(2) ), "right edge is loop" ); 2036 2037 // If I am a shared header (multiple backedges), peel off the many 2038 // backedges into a private merge point and use the merge point as 2039 // the one true backedge. 2040 if( _head->req() > 3 ) { 2041 // Merge the many backedges into a single backedge but leave 2042 // the hottest backedge as separate edge for the following peel. 2043 merge_many_backedges( phase ); 2044 result = true; 2045 } 2046 2047 // If I have one hot backedge, peel off myself loop. 2048 // I better be the outermost loop. 2049 if (_head->req() > 3 && !_irreducible) { 2050 split_outer_loop( phase ); 2051 result = true; 2052 2053 } else if (!_head->is_Loop() && !_irreducible) { 2054 // Make a new LoopNode to replace the old loop head 2055 Node *l = new LoopNode( _head->in(1), _head->in(2) ); 2056 l = igvn.register_new_node_with_optimizer(l, _head); 2057 phase->set_created_loop_node(); 2058 // Go ahead and replace _head 2059 phase->_igvn.replace_node( _head, l ); 2060 _head = l; 2061 phase->set_loop(_head, this); 2062 } 2063 2064 // Now recursively beautify nested loops 2065 if( _child ) result |= _child->beautify_loops( phase ); 2066 if( _next ) result |= _next ->beautify_loops( phase ); 2067 return result; 2068 } 2069 2070 //------------------------------allpaths_check_safepts---------------------------- 2071 // Allpaths backwards scan from loop tail, terminating each path at first safepoint 2072 // encountered. Helper for check_safepts. 2073 void IdealLoopTree::allpaths_check_safepts(VectorSet &visited, Node_List &stack) { 2074 assert(stack.size() == 0, "empty stack"); 2075 stack.push(_tail); 2076 visited.Clear(); 2077 visited.set(_tail->_idx); 2078 while (stack.size() > 0) { 2079 Node* n = stack.pop(); 2080 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2081 // Terminate this path 2082 } else if (n->Opcode() == Op_SafePoint) { 2083 if (_phase->get_loop(n) != this) { 2084 if (_required_safept == NULL) _required_safept = new Node_List(); 2085 _required_safept->push(n); // save the one closest to the tail 2086 } 2087 // Terminate this path 2088 } else { 2089 uint start = n->is_Region() ? 1 : 0; 2090 uint end = n->is_Region() && !n->is_Loop() ? n->req() : start + 1; 2091 for (uint i = start; i < end; i++) { 2092 Node* in = n->in(i); 2093 assert(in->is_CFG(), "must be"); 2094 if (!visited.test_set(in->_idx) && is_member(_phase->get_loop(in))) { 2095 stack.push(in); 2096 } 2097 } 2098 } 2099 } 2100 } 2101 2102 //------------------------------check_safepts---------------------------- 2103 // Given dominators, try to find loops with calls that must always be 2104 // executed (call dominates loop tail). These loops do not need non-call 2105 // safepoints (ncsfpt). 2106 // 2107 // A complication is that a safepoint in a inner loop may be needed 2108 // by an outer loop. In the following, the inner loop sees it has a 2109 // call (block 3) on every path from the head (block 2) to the 2110 // backedge (arc 3->2). So it deletes the ncsfpt (non-call safepoint) 2111 // in block 2, _but_ this leaves the outer loop without a safepoint. 2112 // 2113 // entry 0 2114 // | 2115 // v 2116 // outer 1,2 +->1 2117 // | | 2118 // | v 2119 // | 2<---+ ncsfpt in 2 2120 // |_/|\ | 2121 // | v | 2122 // inner 2,3 / 3 | call in 3 2123 // / | | 2124 // v +--+ 2125 // exit 4 2126 // 2127 // 2128 // This method creates a list (_required_safept) of ncsfpt nodes that must 2129 // be protected is created for each loop. When a ncsfpt maybe deleted, it 2130 // is first looked for in the lists for the outer loops of the current loop. 2131 // 2132 // The insights into the problem: 2133 // A) counted loops are okay 2134 // B) innermost loops are okay (only an inner loop can delete 2135 // a ncsfpt needed by an outer loop) 2136 // C) a loop is immune from an inner loop deleting a safepoint 2137 // if the loop has a call on the idom-path 2138 // D) a loop is also immune if it has a ncsfpt (non-call safepoint) on the 2139 // idom-path that is not in a nested loop 2140 // E) otherwise, an ncsfpt on the idom-path that is nested in an inner 2141 // loop needs to be prevented from deletion by an inner loop 2142 // 2143 // There are two analyses: 2144 // 1) The first, and cheaper one, scans the loop body from 2145 // tail to head following the idom (immediate dominator) 2146 // chain, looking for the cases (C,D,E) above. 2147 // Since inner loops are scanned before outer loops, there is summary 2148 // information about inner loops. Inner loops can be skipped over 2149 // when the tail of an inner loop is encountered. 2150 // 2151 // 2) The second, invoked if the first fails to find a call or ncsfpt on 2152 // the idom path (which is rare), scans all predecessor control paths 2153 // from the tail to the head, terminating a path when a call or sfpt 2154 // is encountered, to find the ncsfpt's that are closest to the tail. 2155 // 2156 void IdealLoopTree::check_safepts(VectorSet &visited, Node_List &stack) { 2157 // Bottom up traversal 2158 IdealLoopTree* ch = _child; 2159 if (_child) _child->check_safepts(visited, stack); 2160 if (_next) _next ->check_safepts(visited, stack); 2161 2162 if (!_head->is_CountedLoop() && !_has_sfpt && _parent != NULL && !_irreducible) { 2163 bool has_call = false; // call on dom-path 2164 bool has_local_ncsfpt = false; // ncsfpt on dom-path at this loop depth 2165 Node* nonlocal_ncsfpt = NULL; // ncsfpt on dom-path at a deeper depth 2166 // Scan the dom-path nodes from tail to head 2167 for (Node* n = tail(); n != _head; n = _phase->idom(n)) { 2168 if (n->is_Call() && n->as_Call()->guaranteed_safepoint()) { 2169 has_call = true; 2170 _has_sfpt = 1; // Then no need for a safept! 2171 break; 2172 } else if (n->Opcode() == Op_SafePoint) { 2173 if (_phase->get_loop(n) == this) { 2174 has_local_ncsfpt = true; 2175 break; 2176 } 2177 if (nonlocal_ncsfpt == NULL) { 2178 nonlocal_ncsfpt = n; // save the one closest to the tail 2179 } 2180 } else { 2181 IdealLoopTree* nlpt = _phase->get_loop(n); 2182 if (this != nlpt) { 2183 // If at an inner loop tail, see if the inner loop has already 2184 // recorded seeing a call on the dom-path (and stop.) If not, 2185 // jump to the head of the inner loop. 2186 assert(is_member(nlpt), "nested loop"); 2187 Node* tail = nlpt->_tail; 2188 if (tail->in(0)->is_If()) tail = tail->in(0); 2189 if (n == tail) { 2190 // If inner loop has call on dom-path, so does outer loop 2191 if (nlpt->_has_sfpt) { 2192 has_call = true; 2193 _has_sfpt = 1; 2194 break; 2195 } 2196 // Skip to head of inner loop 2197 assert(_phase->is_dominator(_head, nlpt->_head), "inner head dominated by outer head"); 2198 n = nlpt->_head; 2199 } 2200 } 2201 } 2202 } 2203 // Record safept's that this loop needs preserved when an 2204 // inner loop attempts to delete it's safepoints. 2205 if (_child != NULL && !has_call && !has_local_ncsfpt) { 2206 if (nonlocal_ncsfpt != NULL) { 2207 if (_required_safept == NULL) _required_safept = new Node_List(); 2208 _required_safept->push(nonlocal_ncsfpt); 2209 } else { 2210 // Failed to find a suitable safept on the dom-path. Now use 2211 // an all paths walk from tail to head, looking for safepoints to preserve. 2212 allpaths_check_safepts(visited, stack); 2213 } 2214 } 2215 } 2216 } 2217 2218 //---------------------------is_deleteable_safept---------------------------- 2219 // Is safept not required by an outer loop? 2220 bool PhaseIdealLoop::is_deleteable_safept(Node* sfpt) { 2221 assert(sfpt->Opcode() == Op_SafePoint, ""); 2222 IdealLoopTree* lp = get_loop(sfpt)->_parent; 2223 while (lp != NULL) { 2224 Node_List* sfpts = lp->_required_safept; 2225 if (sfpts != NULL) { 2226 for (uint i = 0; i < sfpts->size(); i++) { 2227 if (sfpt == sfpts->at(i)) 2228 return false; 2229 } 2230 } 2231 lp = lp->_parent; 2232 } 2233 return true; 2234 } 2235 2236 //---------------------------replace_parallel_iv------------------------------- 2237 // Replace parallel induction variable (parallel to trip counter) 2238 void PhaseIdealLoop::replace_parallel_iv(IdealLoopTree *loop) { 2239 assert(loop->_head->is_CountedLoop(), ""); 2240 CountedLoopNode *cl = loop->_head->as_CountedLoop(); 2241 if (!cl->is_valid_counted_loop()) 2242 return; // skip malformed counted loop 2243 Node *incr = cl->incr(); 2244 if (incr == NULL) 2245 return; // Dead loop? 2246 Node *init = cl->init_trip(); 2247 Node *phi = cl->phi(); 2248 int stride_con = cl->stride_con(); 2249 2250 // Visit all children, looking for Phis 2251 for (DUIterator i = cl->outs(); cl->has_out(i); i++) { 2252 Node *out = cl->out(i); 2253 // Look for other phis (secondary IVs). Skip dead ones 2254 if (!out->is_Phi() || out == phi || !has_node(out)) 2255 continue; 2256 PhiNode* phi2 = out->as_Phi(); 2257 Node *incr2 = phi2->in( LoopNode::LoopBackControl ); 2258 // Look for induction variables of the form: X += constant 2259 if (phi2->region() != loop->_head || 2260 incr2->req() != 3 || 2261 incr2->in(1) != phi2 || 2262 incr2 == incr || 2263 incr2->Opcode() != Op_AddI || 2264 !incr2->in(2)->is_Con()) 2265 continue; 2266 2267 // Check for parallel induction variable (parallel to trip counter) 2268 // via an affine function. In particular, count-down loops with 2269 // count-up array indices are common. We only RCE references off 2270 // the trip-counter, so we need to convert all these to trip-counter 2271 // expressions. 2272 Node *init2 = phi2->in( LoopNode::EntryControl ); 2273 int stride_con2 = incr2->in(2)->get_int(); 2274 2275 // The ratio of the two strides cannot be represented as an int 2276 // if stride_con2 is min_int and stride_con is -1. 2277 if (stride_con2 == min_jint && stride_con == -1) { 2278 continue; 2279 } 2280 2281 // The general case here gets a little tricky. We want to find the 2282 // GCD of all possible parallel IV's and make a new IV using this 2283 // GCD for the loop. Then all possible IVs are simple multiples of 2284 // the GCD. In practice, this will cover very few extra loops. 2285 // Instead we require 'stride_con2' to be a multiple of 'stride_con', 2286 // where +/-1 is the common case, but other integer multiples are 2287 // also easy to handle. 2288 int ratio_con = stride_con2/stride_con; 2289 2290 if ((ratio_con * stride_con) == stride_con2) { // Check for exact 2291 #ifndef PRODUCT 2292 if (TraceLoopOpts) { 2293 tty->print("Parallel IV: %d ", phi2->_idx); 2294 loop->dump_head(); 2295 } 2296 #endif 2297 // Convert to using the trip counter. The parallel induction 2298 // variable differs from the trip counter by a loop-invariant 2299 // amount, the difference between their respective initial values. 2300 // It is scaled by the 'ratio_con'. 2301 Node* ratio = _igvn.intcon(ratio_con); 2302 set_ctrl(ratio, C->root()); 2303 Node* ratio_init = new MulINode(init, ratio); 2304 _igvn.register_new_node_with_optimizer(ratio_init, init); 2305 set_early_ctrl(ratio_init); 2306 Node* diff = new SubINode(init2, ratio_init); 2307 _igvn.register_new_node_with_optimizer(diff, init2); 2308 set_early_ctrl(diff); 2309 Node* ratio_idx = new MulINode(phi, ratio); 2310 _igvn.register_new_node_with_optimizer(ratio_idx, phi); 2311 set_ctrl(ratio_idx, cl); 2312 Node* add = new AddINode(ratio_idx, diff); 2313 _igvn.register_new_node_with_optimizer(add); 2314 set_ctrl(add, cl); 2315 _igvn.replace_node( phi2, add ); 2316 // Sometimes an induction variable is unused 2317 if (add->outcnt() == 0) { 2318 _igvn.remove_dead_node(add); 2319 } 2320 --i; // deleted this phi; rescan starting with next position 2321 continue; 2322 } 2323 } 2324 } 2325 2326 void IdealLoopTree::remove_safepoints(PhaseIdealLoop* phase, bool keep_one) { 2327 Node* keep = NULL; 2328 if (keep_one) { 2329 // Look for a safepoint on the idom-path. 2330 for (Node* i = tail(); i != _head; i = phase->idom(i)) { 2331 if (i->Opcode() == Op_SafePoint && phase->get_loop(i) == this) { 2332 keep = i; 2333 break; // Found one 2334 } 2335 } 2336 } 2337 2338 // Don't remove any safepoints if it is requested to keep a single safepoint and 2339 // no safepoint was found on idom-path. It is not safe to remove any safepoint 2340 // in this case since there's no safepoint dominating all paths in the loop body. 2341 bool prune = !keep_one || keep != NULL; 2342 2343 // Delete other safepoints in this loop. 2344 Node_List* sfpts = _safepts; 2345 if (prune && sfpts != NULL) { 2346 assert(keep == NULL || keep->Opcode() == Op_SafePoint, "not safepoint"); 2347 for (uint i = 0; i < sfpts->size(); i++) { 2348 Node* n = sfpts->at(i); 2349 assert(phase->get_loop(n) == this, ""); 2350 if (n != keep && phase->is_deleteable_safept(n)) { 2351 phase->lazy_replace(n, n->in(TypeFunc::Control)); 2352 } 2353 } 2354 } 2355 } 2356 2357 //------------------------------counted_loop----------------------------------- 2358 // Convert to counted loops where possible 2359 void IdealLoopTree::counted_loop( PhaseIdealLoop *phase ) { 2360 2361 // For grins, set the inner-loop flag here 2362 if (!_child) { 2363 if (_head->is_Loop()) _head->as_Loop()->set_inner_loop(); 2364 } 2365 2366 IdealLoopTree* loop = this; 2367 if (_head->is_CountedLoop() || 2368 phase->is_counted_loop(_head, loop)) { 2369 2370 if (LoopStripMiningIter == 0 || (LoopStripMiningIter > 1 && _child == NULL)) { 2371 // Indicate we do not need a safepoint here 2372 _has_sfpt = 1; 2373 } 2374 2375 // Remove safepoints 2376 bool keep_one_sfpt = !(_has_call || _has_sfpt); 2377 remove_safepoints(phase, keep_one_sfpt); 2378 2379 // Look for induction variables 2380 phase->replace_parallel_iv(this); 2381 2382 } else if (_parent != NULL && !_irreducible) { 2383 // Not a counted loop. Keep one safepoint. 2384 bool keep_one_sfpt = true; 2385 remove_safepoints(phase, keep_one_sfpt); 2386 } 2387 2388 // Recursively 2389 assert(loop->_child != this || (loop->_head->as_Loop()->is_OuterStripMinedLoop() && _head->as_CountedLoop()->is_strip_mined()), "what kind of loop was added?"); 2390 assert(loop->_child != this || (loop->_child->_child == NULL && loop->_child->_next == NULL), "would miss some loops"); 2391 if (loop->_child && loop->_child != this) loop->_child->counted_loop(phase); 2392 if (loop->_next) loop->_next ->counted_loop(phase); 2393 } 2394 2395 #ifndef PRODUCT 2396 //------------------------------dump_head-------------------------------------- 2397 // Dump 1 liner for loop header info 2398 void IdealLoopTree::dump_head( ) const { 2399 for (uint i=0; i<_nest; i++) 2400 tty->print(" "); 2401 tty->print("Loop: N%d/N%d ",_head->_idx,_tail->_idx); 2402 if (_irreducible) tty->print(" IRREDUCIBLE"); 2403 Node* entry = _head->is_Loop() ? _head->as_Loop()->skip_strip_mined(-1)->in(LoopNode::EntryControl) : _head->in(LoopNode::EntryControl); 2404 Node* predicate = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_loop_limit_check); 2405 if (predicate != NULL ) { 2406 tty->print(" limit_check"); 2407 entry = entry->in(0)->in(0); 2408 } 2409 if (UseLoopPredicate) { 2410 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_predicate); 2411 if (entry != NULL) { 2412 tty->print(" predicated"); 2413 entry = PhaseIdealLoop::skip_loop_predicates(entry); 2414 } 2415 } 2416 if (UseProfiledLoopPredicate) { 2417 entry = PhaseIdealLoop::find_predicate_insertion_point(entry, Deoptimization::Reason_profile_predicate); 2418 if (entry != NULL) { 2419 tty->print(" profile_predicated"); 2420 } 2421 } 2422 if (_head->is_CountedLoop()) { 2423 CountedLoopNode *cl = _head->as_CountedLoop(); 2424 tty->print(" counted"); 2425 2426 Node* init_n = cl->init_trip(); 2427 if (init_n != NULL && init_n->is_Con()) 2428 tty->print(" [%d,", cl->init_trip()->get_int()); 2429 else 2430 tty->print(" [int,"); 2431 Node* limit_n = cl->limit(); 2432 if (limit_n != NULL && limit_n->is_Con()) 2433 tty->print("%d),", cl->limit()->get_int()); 2434 else 2435 tty->print("int),"); 2436 int stride_con = cl->stride_con(); 2437 if (stride_con > 0) tty->print("+"); 2438 tty->print("%d", stride_con); 2439 2440 tty->print(" (%0.f iters) ", cl->profile_trip_cnt()); 2441 2442 if (cl->is_pre_loop ()) tty->print(" pre" ); 2443 if (cl->is_main_loop()) tty->print(" main"); 2444 if (cl->is_post_loop()) tty->print(" post"); 2445 if (cl->is_vectorized_loop()) tty->print(" vector"); 2446 if (cl->range_checks_present()) tty->print(" rc "); 2447 if (cl->is_multiversioned()) tty->print(" multi "); 2448 } 2449 if (_has_call) tty->print(" has_call"); 2450 if (_has_sfpt) tty->print(" has_sfpt"); 2451 if (_rce_candidate) tty->print(" rce"); 2452 if (_safepts != NULL && _safepts->size() > 0) { 2453 tty->print(" sfpts={"); _safepts->dump_simple(); tty->print(" }"); 2454 } 2455 if (_required_safept != NULL && _required_safept->size() > 0) { 2456 tty->print(" req={"); _required_safept->dump_simple(); tty->print(" }"); 2457 } 2458 if (Verbose) { 2459 tty->print(" body={"); _body.dump_simple(); tty->print(" }"); 2460 } 2461 if (_head->is_Loop() && _head->as_Loop()->is_strip_mined()) { 2462 tty->print(" strip_mined"); 2463 } 2464 tty->cr(); 2465 } 2466 2467 //------------------------------dump------------------------------------------- 2468 // Dump loops by loop tree 2469 void IdealLoopTree::dump( ) const { 2470 dump_head(); 2471 if (_child) _child->dump(); 2472 if (_next) _next ->dump(); 2473 } 2474 2475 #endif 2476 2477 static void log_loop_tree(IdealLoopTree* root, IdealLoopTree* loop, CompileLog* log) { 2478 if (loop == root) { 2479 if (loop->_child != NULL) { 2480 log->begin_head("loop_tree"); 2481 log->end_head(); 2482 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2483 log->tail("loop_tree"); 2484 assert(loop->_next == NULL, "what?"); 2485 } 2486 } else { 2487 Node* head = loop->_head; 2488 log->begin_head("loop"); 2489 log->print(" idx='%d' ", head->_idx); 2490 if (loop->_irreducible) log->print("irreducible='1' "); 2491 if (head->is_Loop()) { 2492 if (head->as_Loop()->is_inner_loop()) log->print("inner_loop='1' "); 2493 if (head->as_Loop()->is_partial_peel_loop()) log->print("partial_peel_loop='1' "); 2494 } 2495 if (head->is_CountedLoop()) { 2496 CountedLoopNode* cl = head->as_CountedLoop(); 2497 if (cl->is_pre_loop()) log->print("pre_loop='%d' ", cl->main_idx()); 2498 if (cl->is_main_loop()) log->print("main_loop='%d' ", cl->_idx); 2499 if (cl->is_post_loop()) log->print("post_loop='%d' ", cl->main_idx()); 2500 } 2501 log->end_head(); 2502 if( loop->_child ) log_loop_tree(root, loop->_child, log); 2503 log->tail("loop"); 2504 if( loop->_next ) log_loop_tree(root, loop->_next, log); 2505 } 2506 } 2507 2508 //---------------------collect_potentially_useful_predicates----------------------- 2509 // Helper function to collect potentially useful predicates to prevent them from 2510 // being eliminated by PhaseIdealLoop::eliminate_useless_predicates 2511 void PhaseIdealLoop::collect_potentially_useful_predicates( 2512 IdealLoopTree * loop, Unique_Node_List &useful_predicates) { 2513 if (loop->_child) { // child 2514 collect_potentially_useful_predicates(loop->_child, useful_predicates); 2515 } 2516 2517 // self (only loops that we can apply loop predication may use their predicates) 2518 if (loop->_head->is_Loop() && 2519 !loop->_irreducible && 2520 !loop->tail()->is_top()) { 2521 LoopNode* lpn = loop->_head->as_Loop(); 2522 Node* entry = lpn->in(LoopNode::EntryControl); 2523 Node* predicate_proj = find_predicate(entry); // loop_limit_check first 2524 if (predicate_proj != NULL ) { // right pattern that can be used by loop predication 2525 assert(entry->in(0)->in(1)->in(1)->Opcode() == Op_Opaque1, "must be"); 2526 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2527 entry = skip_loop_predicates(entry); 2528 } 2529 predicate_proj = find_predicate(entry); // Predicate 2530 if (predicate_proj != NULL ) { 2531 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2532 entry = skip_loop_predicates(entry); 2533 } 2534 if (UseProfiledLoopPredicate) { 2535 predicate_proj = find_predicate(entry); // Predicate 2536 if (predicate_proj != NULL ) { 2537 useful_predicates.push(entry->in(0)->in(1)->in(1)); // good one 2538 } 2539 } 2540 } 2541 2542 if (loop->_next) { // sibling 2543 collect_potentially_useful_predicates(loop->_next, useful_predicates); 2544 } 2545 } 2546 2547 //------------------------eliminate_useless_predicates----------------------------- 2548 // Eliminate all inserted predicates if they could not be used by loop predication. 2549 // Note: it will also eliminates loop limits check predicate since it also uses 2550 // Opaque1 node (see Parse::add_predicate()). 2551 void PhaseIdealLoop::eliminate_useless_predicates() { 2552 if (C->predicate_count() == 0) 2553 return; // no predicate left 2554 2555 Unique_Node_List useful_predicates; // to store useful predicates 2556 if (C->has_loops()) { 2557 collect_potentially_useful_predicates(_ltree_root->_child, useful_predicates); 2558 } 2559 2560 for (int i = C->predicate_count(); i > 0; i--) { 2561 Node * n = C->predicate_opaque1_node(i-1); 2562 assert(n->Opcode() == Op_Opaque1, "must be"); 2563 if (!useful_predicates.member(n)) { // not in the useful list 2564 _igvn.replace_node(n, n->in(1)); 2565 } 2566 } 2567 } 2568 2569 //------------------------process_expensive_nodes----------------------------- 2570 // Expensive nodes have their control input set to prevent the GVN 2571 // from commoning them and as a result forcing the resulting node to 2572 // be in a more frequent path. Use CFG information here, to change the 2573 // control inputs so that some expensive nodes can be commoned while 2574 // not executed more frequently. 2575 bool PhaseIdealLoop::process_expensive_nodes() { 2576 assert(OptimizeExpensiveOps, "optimization off?"); 2577 2578 // Sort nodes to bring similar nodes together 2579 C->sort_expensive_nodes(); 2580 2581 bool progress = false; 2582 2583 for (int i = 0; i < C->expensive_count(); ) { 2584 Node* n = C->expensive_node(i); 2585 int start = i; 2586 // Find nodes similar to n 2587 i++; 2588 for (; i < C->expensive_count() && Compile::cmp_expensive_nodes(n, C->expensive_node(i)) == 0; i++); 2589 int end = i; 2590 // And compare them two by two 2591 for (int j = start; j < end; j++) { 2592 Node* n1 = C->expensive_node(j); 2593 if (is_node_unreachable(n1)) { 2594 continue; 2595 } 2596 for (int k = j+1; k < end; k++) { 2597 Node* n2 = C->expensive_node(k); 2598 if (is_node_unreachable(n2)) { 2599 continue; 2600 } 2601 2602 assert(n1 != n2, "should be pair of nodes"); 2603 2604 Node* c1 = n1->in(0); 2605 Node* c2 = n2->in(0); 2606 2607 Node* parent_c1 = c1; 2608 Node* parent_c2 = c2; 2609 2610 // The call to get_early_ctrl_for_expensive() moves the 2611 // expensive nodes up but stops at loops that are in a if 2612 // branch. See whether we can exit the loop and move above the 2613 // If. 2614 if (c1->is_Loop()) { 2615 parent_c1 = c1->in(1); 2616 } 2617 if (c2->is_Loop()) { 2618 parent_c2 = c2->in(1); 2619 } 2620 2621 if (parent_c1 == parent_c2) { 2622 _igvn._worklist.push(n1); 2623 _igvn._worklist.push(n2); 2624 continue; 2625 } 2626 2627 // Look for identical expensive node up the dominator chain. 2628 if (is_dominator(c1, c2)) { 2629 c2 = c1; 2630 } else if (is_dominator(c2, c1)) { 2631 c1 = c2; 2632 } else if (parent_c1->is_Proj() && parent_c1->in(0)->is_If() && 2633 parent_c2->is_Proj() && parent_c1->in(0) == parent_c2->in(0)) { 2634 // Both branches have the same expensive node so move it up 2635 // before the if. 2636 c1 = c2 = idom(parent_c1->in(0)); 2637 } 2638 // Do the actual moves 2639 if (n1->in(0) != c1) { 2640 _igvn.hash_delete(n1); 2641 n1->set_req(0, c1); 2642 _igvn.hash_insert(n1); 2643 _igvn._worklist.push(n1); 2644 progress = true; 2645 } 2646 if (n2->in(0) != c2) { 2647 _igvn.hash_delete(n2); 2648 n2->set_req(0, c2); 2649 _igvn.hash_insert(n2); 2650 _igvn._worklist.push(n2); 2651 progress = true; 2652 } 2653 } 2654 } 2655 } 2656 2657 return progress; 2658 } 2659 2660 2661 //============================================================================= 2662 //----------------------------build_and_optimize------------------------------- 2663 // Create a PhaseLoop. Build the ideal Loop tree. Map each Ideal Node to 2664 // its corresponding LoopNode. If 'optimize' is true, do some loop cleanups. 2665 void PhaseIdealLoop::build_and_optimize(LoopOptsMode mode) { 2666 bool do_split_ifs = (mode == LoopOptsDefault || mode == LoopOptsLastRound); 2667 bool skip_loop_opts = (mode == LoopOptsNone); 2668 2669 ResourceMark rm; 2670 2671 int old_progress = C->major_progress(); 2672 uint orig_worklist_size = _igvn._worklist.size(); 2673 2674 // Reset major-progress flag for the driver's heuristics 2675 C->clear_major_progress(); 2676 2677 #ifndef PRODUCT 2678 // Capture for later assert 2679 uint unique = C->unique(); 2680 _loop_invokes++; 2681 _loop_work += unique; 2682 #endif 2683 2684 // True if the method has at least 1 irreducible loop 2685 _has_irreducible_loops = false; 2686 2687 _created_loop_node = false; 2688 2689 Arena *a = Thread::current()->resource_area(); 2690 VectorSet visited(a); 2691 // Pre-grow the mapping from Nodes to IdealLoopTrees. 2692 _nodes.map(C->unique(), NULL); 2693 memset(_nodes.adr(), 0, wordSize * C->unique()); 2694 2695 // Pre-build the top-level outermost loop tree entry 2696 _ltree_root = new IdealLoopTree( this, C->root(), C->root() ); 2697 // Do not need a safepoint at the top level 2698 _ltree_root->_has_sfpt = 1; 2699 2700 // Initialize Dominators. 2701 // Checked in clone_loop_predicate() during beautify_loops(). 2702 _idom_size = 0; 2703 _idom = NULL; 2704 _dom_depth = NULL; 2705 _dom_stk = NULL; 2706 2707 // Empty pre-order array 2708 allocate_preorders(); 2709 2710 // Build a loop tree on the fly. Build a mapping from CFG nodes to 2711 // IdealLoopTree entries. Data nodes are NOT walked. 2712 build_loop_tree(); 2713 // Check for bailout, and return 2714 if (C->failing()) { 2715 return; 2716 } 2717 2718 // No loops after all 2719 if( !_ltree_root->_child && !_verify_only ) C->set_has_loops(false); 2720 2721 // There should always be an outer loop containing the Root and Return nodes. 2722 // If not, we have a degenerate empty program. Bail out in this case. 2723 if (!has_node(C->root())) { 2724 if (!_verify_only) { 2725 C->clear_major_progress(); 2726 C->record_method_not_compilable("empty program detected during loop optimization"); 2727 } 2728 return; 2729 } 2730 2731 // Nothing to do, so get out 2732 bool stop_early = !C->has_loops() && !skip_loop_opts && !do_split_ifs && !_verify_me && !_verify_only; 2733 bool do_expensive_nodes = C->should_optimize_expensive_nodes(_igvn); 2734 if (stop_early && !do_expensive_nodes) { 2735 _igvn.optimize(); // Cleanup NeverBranches 2736 return; 2737 } 2738 2739 // Set loop nesting depth 2740 _ltree_root->set_nest( 0 ); 2741 2742 // Split shared headers and insert loop landing pads. 2743 // Do not bother doing this on the Root loop of course. 2744 if( !_verify_me && !_verify_only && _ltree_root->_child ) { 2745 C->print_method(PHASE_BEFORE_BEAUTIFY_LOOPS, 3); 2746 if( _ltree_root->_child->beautify_loops( this ) ) { 2747 // Re-build loop tree! 2748 _ltree_root->_child = NULL; 2749 _nodes.clear(); 2750 reallocate_preorders(); 2751 build_loop_tree(); 2752 // Check for bailout, and return 2753 if (C->failing()) { 2754 return; 2755 } 2756 // Reset loop nesting depth 2757 _ltree_root->set_nest( 0 ); 2758 2759 C->print_method(PHASE_AFTER_BEAUTIFY_LOOPS, 3); 2760 } 2761 } 2762 2763 // Build Dominators for elision of NULL checks & loop finding. 2764 // Since nodes do not have a slot for immediate dominator, make 2765 // a persistent side array for that info indexed on node->_idx. 2766 _idom_size = C->unique(); 2767 _idom = NEW_RESOURCE_ARRAY( Node*, _idom_size ); 2768 _dom_depth = NEW_RESOURCE_ARRAY( uint, _idom_size ); 2769 _dom_stk = NULL; // Allocated on demand in recompute_dom_depth 2770 memset( _dom_depth, 0, _idom_size * sizeof(uint) ); 2771 2772 Dominators(); 2773 2774 if (!_verify_only) { 2775 // As a side effect, Dominators removed any unreachable CFG paths 2776 // into RegionNodes. It doesn't do this test against Root, so 2777 // we do it here. 2778 for( uint i = 1; i < C->root()->req(); i++ ) { 2779 if( !_nodes[C->root()->in(i)->_idx] ) { // Dead path into Root? 2780 _igvn.delete_input_of(C->root(), i); 2781 i--; // Rerun same iteration on compressed edges 2782 } 2783 } 2784 2785 // Given dominators, try to find inner loops with calls that must 2786 // always be executed (call dominates loop tail). These loops do 2787 // not need a separate safepoint. 2788 Node_List cisstack(a); 2789 _ltree_root->check_safepts(visited, cisstack); 2790 } 2791 2792 // Walk the DATA nodes and place into loops. Find earliest control 2793 // node. For CFG nodes, the _nodes array starts out and remains 2794 // holding the associated IdealLoopTree pointer. For DATA nodes, the 2795 // _nodes array holds the earliest legal controlling CFG node. 2796 2797 // Allocate stack with enough space to avoid frequent realloc 2798 int stack_size = (C->live_nodes() >> 1) + 16; // (live_nodes>>1)+16 from Java2D stats 2799 Node_Stack nstack( a, stack_size ); 2800 2801 visited.Clear(); 2802 Node_List worklist(a); 2803 // Don't need C->root() on worklist since 2804 // it will be processed among C->top() inputs 2805 worklist.push( C->top() ); 2806 visited.set( C->top()->_idx ); // Set C->top() as visited now 2807 build_loop_early( visited, worklist, nstack ); 2808 2809 // Given early legal placement, try finding counted loops. This placement 2810 // is good enough to discover most loop invariants. 2811 if( !_verify_me && !_verify_only ) 2812 _ltree_root->counted_loop( this ); 2813 2814 // Find latest loop placement. Find ideal loop placement. 2815 visited.Clear(); 2816 init_dom_lca_tags(); 2817 // Need C->root() on worklist when processing outs 2818 worklist.push( C->root() ); 2819 NOT_PRODUCT( C->verify_graph_edges(); ) 2820 worklist.push( C->top() ); 2821 build_loop_late( visited, worklist, nstack ); 2822 2823 if (_verify_only) { 2824 // restore major progress flag 2825 for (int i = 0; i < old_progress; i++) 2826 C->set_major_progress(); 2827 assert(C->unique() == unique, "verification mode made Nodes? ? ?"); 2828 assert(_igvn._worklist.size() == orig_worklist_size, "shouldn't push anything"); 2829 return; 2830 } 2831 2832 // clear out the dead code after build_loop_late 2833 while (_deadlist.size()) { 2834 _igvn.remove_globally_dead_node(_deadlist.pop()); 2835 } 2836 2837 if (stop_early) { 2838 assert(do_expensive_nodes, "why are we here?"); 2839 if (process_expensive_nodes()) { 2840 // If we made some progress when processing expensive nodes then 2841 // the IGVN may modify the graph in a way that will allow us to 2842 // make some more progress: we need to try processing expensive 2843 // nodes again. 2844 C->set_major_progress(); 2845 } 2846 _igvn.optimize(); 2847 return; 2848 } 2849 2850 // Some parser-inserted loop predicates could never be used by loop 2851 // predication or they were moved away from loop during some optimizations. 2852 // For example, peeling. Eliminate them before next loop optimizations. 2853 eliminate_useless_predicates(); 2854 2855 #ifndef PRODUCT 2856 C->verify_graph_edges(); 2857 if (_verify_me) { // Nested verify pass? 2858 // Check to see if the verify mode is broken 2859 assert(C->unique() == unique, "non-optimize mode made Nodes? ? ?"); 2860 return; 2861 } 2862 if(VerifyLoopOptimizations) verify(); 2863 if(TraceLoopOpts && C->has_loops()) { 2864 _ltree_root->dump(); 2865 } 2866 #endif 2867 2868 if (skip_loop_opts) { 2869 // restore major progress flag 2870 for (int i = 0; i < old_progress; i++) { 2871 C->set_major_progress(); 2872 } 2873 2874 // Cleanup any modified bits 2875 _igvn.optimize(); 2876 2877 if (C->log() != NULL) { 2878 log_loop_tree(_ltree_root, _ltree_root, C->log()); 2879 } 2880 return; 2881 } 2882 2883 if (ReassociateInvariants) { 2884 // Reassociate invariants and prep for split_thru_phi 2885 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2886 IdealLoopTree* lpt = iter.current(); 2887 bool is_counted = lpt->is_counted(); 2888 if (!is_counted || !lpt->is_inner()) continue; 2889 2890 // check for vectorized loops, any reassociation of invariants was already done 2891 if (is_counted && lpt->_head->as_CountedLoop()->do_unroll_only()) continue; 2892 2893 lpt->reassociate_invariants(this); 2894 2895 // Because RCE opportunities can be masked by split_thru_phi, 2896 // look for RCE candidates and inhibit split_thru_phi 2897 // on just their loop-phi's for this pass of loop opts 2898 if (SplitIfBlocks && do_split_ifs) { 2899 if (lpt->policy_range_check(this)) { 2900 lpt->_rce_candidate = 1; // = true 2901 } 2902 } 2903 } 2904 } 2905 2906 // Check for aggressive application of split-if and other transforms 2907 // that require basic-block info (like cloning through Phi's) 2908 if( SplitIfBlocks && do_split_ifs ) { 2909 visited.Clear(); 2910 split_if_with_blocks( visited, nstack, mode == LoopOptsLastRound ); 2911 NOT_PRODUCT( if( VerifyLoopOptimizations ) verify(); ); 2912 if (mode == LoopOptsLastRound) { 2913 C->set_major_progress(); 2914 } 2915 } 2916 2917 if (!C->major_progress() && do_expensive_nodes && process_expensive_nodes()) { 2918 C->set_major_progress(); 2919 } 2920 2921 // Perform loop predication before iteration splitting 2922 if (C->has_loops() && !C->major_progress() && (C->predicate_count() > 0)) { 2923 _ltree_root->_child->loop_predication(this); 2924 } 2925 2926 if (OptimizeFill && UseLoopPredicate && C->has_loops() && !C->major_progress()) { 2927 if (do_intrinsify_fill()) { 2928 C->set_major_progress(); 2929 } 2930 } 2931 2932 // Perform iteration-splitting on inner loops. Split iterations to avoid 2933 // range checks or one-shot null checks. 2934 2935 // If split-if's didn't hack the graph too bad (no CFG changes) 2936 // then do loop opts. 2937 if (C->has_loops() && !C->major_progress()) { 2938 memset( worklist.adr(), 0, worklist.Size()*sizeof(Node*) ); 2939 _ltree_root->_child->iteration_split( this, worklist ); 2940 // No verify after peeling! GCM has hoisted code out of the loop. 2941 // After peeling, the hoisted code could sink inside the peeled area. 2942 // The peeling code does not try to recompute the best location for 2943 // all the code before the peeled area, so the verify pass will always 2944 // complain about it. 2945 } 2946 // Do verify graph edges in any case 2947 NOT_PRODUCT( C->verify_graph_edges(); ); 2948 2949 if (!do_split_ifs) { 2950 // We saw major progress in Split-If to get here. We forced a 2951 // pass with unrolling and not split-if, however more split-if's 2952 // might make progress. If the unrolling didn't make progress 2953 // then the major-progress flag got cleared and we won't try 2954 // another round of Split-If. In particular the ever-common 2955 // instance-of/check-cast pattern requires at least 2 rounds of 2956 // Split-If to clear out. 2957 C->set_major_progress(); 2958 } 2959 2960 // Repeat loop optimizations if new loops were seen 2961 if (created_loop_node()) { 2962 C->set_major_progress(); 2963 } 2964 2965 // Keep loop predicates and perform optimizations with them 2966 // until no more loop optimizations could be done. 2967 // After that switch predicates off and do more loop optimizations. 2968 if (!C->major_progress() && (C->predicate_count() > 0)) { 2969 C->cleanup_loop_predicates(_igvn); 2970 if (TraceLoopOpts) { 2971 tty->print_cr("PredicatesOff"); 2972 } 2973 C->set_major_progress(); 2974 } 2975 2976 // Convert scalar to superword operations at the end of all loop opts. 2977 if (UseSuperWord && C->has_loops() && !C->major_progress()) { 2978 // SuperWord transform 2979 SuperWord sw(this); 2980 for (LoopTreeIterator iter(_ltree_root); !iter.done(); iter.next()) { 2981 IdealLoopTree* lpt = iter.current(); 2982 if (lpt->is_counted()) { 2983 CountedLoopNode *cl = lpt->_head->as_CountedLoop(); 2984 2985 if (PostLoopMultiversioning && cl->is_rce_post_loop() && !cl->is_vectorized_loop()) { 2986 // Check that the rce'd post loop is encountered first, multiversion after all 2987 // major main loop optimization are concluded 2988 if (!C->major_progress()) { 2989 IdealLoopTree *lpt_next = lpt->_next; 2990 if (lpt_next && lpt_next->is_counted()) { 2991 CountedLoopNode *cl = lpt_next->_head->as_CountedLoop(); 2992 has_range_checks(lpt_next); 2993 if (cl->is_post_loop() && cl->range_checks_present()) { 2994 if (!cl->is_multiversioned()) { 2995 if (multi_version_post_loops(lpt, lpt_next) == false) { 2996 // Cause the rce loop to be optimized away if we fail 2997 cl->mark_is_multiversioned(); 2998 cl->set_slp_max_unroll(0); 2999 poison_rce_post_loop(lpt); 3000 } 3001 } 3002 } 3003 } 3004 sw.transform_loop(lpt, true); 3005 } 3006 } else if (cl->is_main_loop()) { 3007 sw.transform_loop(lpt, true); 3008 } 3009 } 3010 } 3011 } 3012 3013 // Cleanup any modified bits 3014 _igvn.optimize(); 3015 3016 // disable assert until issue with split_flow_path is resolved (6742111) 3017 // assert(!_has_irreducible_loops || C->parsed_irreducible_loop() || C->is_osr_compilation(), 3018 // "shouldn't introduce irreducible loops"); 3019 3020 if (C->log() != NULL) { 3021 log_loop_tree(_ltree_root, _ltree_root, C->log()); 3022 } 3023 } 3024 3025 #ifndef PRODUCT 3026 //------------------------------print_statistics------------------------------- 3027 int PhaseIdealLoop::_loop_invokes=0;// Count of PhaseIdealLoop invokes 3028 int PhaseIdealLoop::_loop_work=0; // Sum of PhaseIdealLoop x unique 3029 void PhaseIdealLoop::print_statistics() { 3030 tty->print_cr("PhaseIdealLoop=%d, sum _unique=%d", _loop_invokes, _loop_work); 3031 } 3032 3033 //------------------------------verify----------------------------------------- 3034 // Build a verify-only PhaseIdealLoop, and see that it agrees with me. 3035 static int fail; // debug only, so its multi-thread dont care 3036 void PhaseIdealLoop::verify() const { 3037 int old_progress = C->major_progress(); 3038 ResourceMark rm; 3039 PhaseIdealLoop loop_verify( _igvn, this ); 3040 VectorSet visited(Thread::current()->resource_area()); 3041 3042 fail = 0; 3043 verify_compare( C->root(), &loop_verify, visited ); 3044 assert( fail == 0, "verify loops failed" ); 3045 // Verify loop structure is the same 3046 _ltree_root->verify_tree(loop_verify._ltree_root, NULL); 3047 // Reset major-progress. It was cleared by creating a verify version of 3048 // PhaseIdealLoop. 3049 for( int i=0; i<old_progress; i++ ) 3050 C->set_major_progress(); 3051 } 3052 3053 //------------------------------verify_compare--------------------------------- 3054 // Make sure me and the given PhaseIdealLoop agree on key data structures 3055 void PhaseIdealLoop::verify_compare( Node *n, const PhaseIdealLoop *loop_verify, VectorSet &visited ) const { 3056 if( !n ) return; 3057 if( visited.test_set( n->_idx ) ) return; 3058 if( !_nodes[n->_idx] ) { // Unreachable 3059 assert( !loop_verify->_nodes[n->_idx], "both should be unreachable" ); 3060 return; 3061 } 3062 3063 uint i; 3064 for( i = 0; i < n->req(); i++ ) 3065 verify_compare( n->in(i), loop_verify, visited ); 3066 3067 // Check the '_nodes' block/loop structure 3068 i = n->_idx; 3069 if( has_ctrl(n) ) { // We have control; verify has loop or ctrl 3070 if( _nodes[i] != loop_verify->_nodes[i] && 3071 get_ctrl_no_update(n) != loop_verify->get_ctrl_no_update(n) ) { 3072 tty->print("Mismatched control setting for: "); 3073 n->dump(); 3074 if( fail++ > 10 ) return; 3075 Node *c = get_ctrl_no_update(n); 3076 tty->print("We have it as: "); 3077 if( c->in(0) ) c->dump(); 3078 else tty->print_cr("N%d",c->_idx); 3079 tty->print("Verify thinks: "); 3080 if( loop_verify->has_ctrl(n) ) 3081 loop_verify->get_ctrl_no_update(n)->dump(); 3082 else 3083 loop_verify->get_loop_idx(n)->dump(); 3084 tty->cr(); 3085 } 3086 } else { // We have a loop 3087 IdealLoopTree *us = get_loop_idx(n); 3088 if( loop_verify->has_ctrl(n) ) { 3089 tty->print("Mismatched loop setting for: "); 3090 n->dump(); 3091 if( fail++ > 10 ) return; 3092 tty->print("We have it as: "); 3093 us->dump(); 3094 tty->print("Verify thinks: "); 3095 loop_verify->get_ctrl_no_update(n)->dump(); 3096 tty->cr(); 3097 } else if (!C->major_progress()) { 3098 // Loop selection can be messed up if we did a major progress 3099 // operation, like split-if. Do not verify in that case. 3100 IdealLoopTree *them = loop_verify->get_loop_idx(n); 3101 if( us->_head != them->_head || us->_tail != them->_tail ) { 3102 tty->print("Unequals loops for: "); 3103 n->dump(); 3104 if( fail++ > 10 ) return; 3105 tty->print("We have it as: "); 3106 us->dump(); 3107 tty->print("Verify thinks: "); 3108 them->dump(); 3109 tty->cr(); 3110 } 3111 } 3112 } 3113 3114 // Check for immediate dominators being equal 3115 if( i >= _idom_size ) { 3116 if( !n->is_CFG() ) return; 3117 tty->print("CFG Node with no idom: "); 3118 n->dump(); 3119 return; 3120 } 3121 if( !n->is_CFG() ) return; 3122 if( n == C->root() ) return; // No IDOM here 3123 3124 assert(n->_idx == i, "sanity"); 3125 Node *id = idom_no_update(n); 3126 if( id != loop_verify->idom_no_update(n) ) { 3127 tty->print("Unequals idoms for: "); 3128 n->dump(); 3129 if( fail++ > 10 ) return; 3130 tty->print("We have it as: "); 3131 id->dump(); 3132 tty->print("Verify thinks: "); 3133 loop_verify->idom_no_update(n)->dump(); 3134 tty->cr(); 3135 } 3136 3137 } 3138 3139 //------------------------------verify_tree------------------------------------ 3140 // Verify that tree structures match. Because the CFG can change, siblings 3141 // within the loop tree can be reordered. We attempt to deal with that by 3142 // reordering the verify's loop tree if possible. 3143 void IdealLoopTree::verify_tree(IdealLoopTree *loop, const IdealLoopTree *parent) const { 3144 assert( _parent == parent, "Badly formed loop tree" ); 3145 3146 // Siblings not in same order? Attempt to re-order. 3147 if( _head != loop->_head ) { 3148 // Find _next pointer to update 3149 IdealLoopTree **pp = &loop->_parent->_child; 3150 while( *pp != loop ) 3151 pp = &((*pp)->_next); 3152 // Find proper sibling to be next 3153 IdealLoopTree **nn = &loop->_next; 3154 while( (*nn) && (*nn)->_head != _head ) 3155 nn = &((*nn)->_next); 3156 3157 // Check for no match. 3158 if( !(*nn) ) { 3159 // Annoyingly, irreducible loops can pick different headers 3160 // after a major_progress operation, so the rest of the loop 3161 // tree cannot be matched. 3162 if (_irreducible && Compile::current()->major_progress()) return; 3163 assert( 0, "failed to match loop tree" ); 3164 } 3165 3166 // Move (*nn) to (*pp) 3167 IdealLoopTree *hit = *nn; 3168 *nn = hit->_next; 3169 hit->_next = loop; 3170 *pp = loop; 3171 loop = hit; 3172 // Now try again to verify 3173 } 3174 3175 assert( _head == loop->_head , "mismatched loop head" ); 3176 Node *tail = _tail; // Inline a non-updating version of 3177 while( !tail->in(0) ) // the 'tail()' call. 3178 tail = tail->in(1); 3179 assert( tail == loop->_tail, "mismatched loop tail" ); 3180 3181 // Counted loops that are guarded should be able to find their guards 3182 if( _head->is_CountedLoop() && _head->as_CountedLoop()->is_main_loop() ) { 3183 CountedLoopNode *cl = _head->as_CountedLoop(); 3184 Node *init = cl->init_trip(); 3185 Node *ctrl = cl->in(LoopNode::EntryControl); 3186 assert( ctrl->Opcode() == Op_IfTrue || ctrl->Opcode() == Op_IfFalse, "" ); 3187 Node *iff = ctrl->in(0); 3188 assert( iff->Opcode() == Op_If, "" ); 3189 Node *bol = iff->in(1); 3190 assert( bol->Opcode() == Op_Bool, "" ); 3191 Node *cmp = bol->in(1); 3192 assert( cmp->Opcode() == Op_CmpI, "" ); 3193 Node *add = cmp->in(1); 3194 Node *opaq; 3195 if( add->Opcode() == Op_Opaque1 ) { 3196 opaq = add; 3197 } else { 3198 assert( add->Opcode() == Op_AddI || add->Opcode() == Op_ConI , "" ); 3199 assert( add == init, "" ); 3200 opaq = cmp->in(2); 3201 } 3202 assert( opaq->Opcode() == Op_Opaque1, "" ); 3203 3204 } 3205 3206 if (_child != NULL) _child->verify_tree(loop->_child, this); 3207 if (_next != NULL) _next ->verify_tree(loop->_next, parent); 3208 // Innermost loops need to verify loop bodies, 3209 // but only if no 'major_progress' 3210 int fail = 0; 3211 if (!Compile::current()->major_progress() && _child == NULL) { 3212 for( uint i = 0; i < _body.size(); i++ ) { 3213 Node *n = _body.at(i); 3214 if (n->outcnt() == 0) continue; // Ignore dead 3215 uint j; 3216 for( j = 0; j < loop->_body.size(); j++ ) 3217 if( loop->_body.at(j) == n ) 3218 break; 3219 if( j == loop->_body.size() ) { // Not found in loop body 3220 // Last ditch effort to avoid assertion: Its possible that we 3221 // have some users (so outcnt not zero) but are still dead. 3222 // Try to find from root. 3223 if (Compile::current()->root()->find(n->_idx)) { 3224 fail++; 3225 tty->print("We have that verify does not: "); 3226 n->dump(); 3227 } 3228 } 3229 } 3230 for( uint i2 = 0; i2 < loop->_body.size(); i2++ ) { 3231 Node *n = loop->_body.at(i2); 3232 if (n->outcnt() == 0) continue; // Ignore dead 3233 uint j; 3234 for( j = 0; j < _body.size(); j++ ) 3235 if( _body.at(j) == n ) 3236 break; 3237 if( j == _body.size() ) { // Not found in loop body 3238 // Last ditch effort to avoid assertion: Its possible that we 3239 // have some users (so outcnt not zero) but are still dead. 3240 // Try to find from root. 3241 if (Compile::current()->root()->find(n->_idx)) { 3242 fail++; 3243 tty->print("Verify has that we do not: "); 3244 n->dump(); 3245 } 3246 } 3247 } 3248 assert( !fail, "loop body mismatch" ); 3249 } 3250 } 3251 3252 #endif 3253 3254 //------------------------------set_idom--------------------------------------- 3255 void PhaseIdealLoop::set_idom(Node* d, Node* n, uint dom_depth) { 3256 uint idx = d->_idx; 3257 if (idx >= _idom_size) { 3258 uint newsize = _idom_size<<1; 3259 while( idx >= newsize ) { 3260 newsize <<= 1; 3261 } 3262 _idom = REALLOC_RESOURCE_ARRAY( Node*, _idom,_idom_size,newsize); 3263 _dom_depth = REALLOC_RESOURCE_ARRAY( uint, _dom_depth,_idom_size,newsize); 3264 memset( _dom_depth + _idom_size, 0, (newsize - _idom_size) * sizeof(uint) ); 3265 _idom_size = newsize; 3266 } 3267 _idom[idx] = n; 3268 _dom_depth[idx] = dom_depth; 3269 } 3270 3271 //------------------------------recompute_dom_depth--------------------------------------- 3272 // The dominator tree is constructed with only parent pointers. 3273 // This recomputes the depth in the tree by first tagging all 3274 // nodes as "no depth yet" marker. The next pass then runs up 3275 // the dom tree from each node marked "no depth yet", and computes 3276 // the depth on the way back down. 3277 void PhaseIdealLoop::recompute_dom_depth() { 3278 uint no_depth_marker = C->unique(); 3279 uint i; 3280 // Initialize depth to "no depth yet" and realize all lazy updates 3281 for (i = 0; i < _idom_size; i++) { 3282 // Only indices with a _dom_depth has a Node* or NULL (otherwise uninitalized). 3283 if (_dom_depth[i] > 0 && _idom[i] != NULL) { 3284 _dom_depth[i] = no_depth_marker; 3285 3286 // heal _idom if it has a fwd mapping in _nodes 3287 if (_idom[i]->in(0) == NULL) { 3288 idom(i); 3289 } 3290 } 3291 } 3292 if (_dom_stk == NULL) { 3293 uint init_size = C->live_nodes() / 100; // Guess that 1/100 is a reasonable initial size. 3294 if (init_size < 10) init_size = 10; 3295 _dom_stk = new GrowableArray<uint>(init_size); 3296 } 3297 // Compute new depth for each node. 3298 for (i = 0; i < _idom_size; i++) { 3299 uint j = i; 3300 // Run up the dom tree to find a node with a depth 3301 while (_dom_depth[j] == no_depth_marker) { 3302 _dom_stk->push(j); 3303 j = _idom[j]->_idx; 3304 } 3305 // Compute the depth on the way back down this tree branch 3306 uint dd = _dom_depth[j] + 1; 3307 while (_dom_stk->length() > 0) { 3308 uint j = _dom_stk->pop(); 3309 _dom_depth[j] = dd; 3310 dd++; 3311 } 3312 } 3313 } 3314 3315 //------------------------------sort------------------------------------------- 3316 // Insert 'loop' into the existing loop tree. 'innermost' is a leaf of the 3317 // loop tree, not the root. 3318 IdealLoopTree *PhaseIdealLoop::sort( IdealLoopTree *loop, IdealLoopTree *innermost ) { 3319 if( !innermost ) return loop; // New innermost loop 3320 3321 int loop_preorder = get_preorder(loop->_head); // Cache pre-order number 3322 assert( loop_preorder, "not yet post-walked loop" ); 3323 IdealLoopTree **pp = &innermost; // Pointer to previous next-pointer 3324 IdealLoopTree *l = *pp; // Do I go before or after 'l'? 3325 3326 // Insert at start of list 3327 while( l ) { // Insertion sort based on pre-order 3328 if( l == loop ) return innermost; // Already on list! 3329 int l_preorder = get_preorder(l->_head); // Cache pre-order number 3330 assert( l_preorder, "not yet post-walked l" ); 3331 // Check header pre-order number to figure proper nesting 3332 if( loop_preorder > l_preorder ) 3333 break; // End of insertion 3334 // If headers tie (e.g., shared headers) check tail pre-order numbers. 3335 // Since I split shared headers, you'd think this could not happen. 3336 // BUT: I must first do the preorder numbering before I can discover I 3337 // have shared headers, so the split headers all get the same preorder 3338 // number as the RegionNode they split from. 3339 if( loop_preorder == l_preorder && 3340 get_preorder(loop->_tail) < get_preorder(l->_tail) ) 3341 break; // Also check for shared headers (same pre#) 3342 pp = &l->_parent; // Chain up list 3343 l = *pp; 3344 } 3345 // Link into list 3346 // Point predecessor to me 3347 *pp = loop; 3348 // Point me to successor 3349 IdealLoopTree *p = loop->_parent; 3350 loop->_parent = l; // Point me to successor 3351 if( p ) sort( p, innermost ); // Insert my parents into list as well 3352 return innermost; 3353 } 3354 3355 //------------------------------build_loop_tree-------------------------------- 3356 // I use a modified Vick/Tarjan algorithm. I need pre- and a post- visit 3357 // bits. The _nodes[] array is mapped by Node index and holds a NULL for 3358 // not-yet-pre-walked, pre-order # for pre-but-not-post-walked and holds the 3359 // tightest enclosing IdealLoopTree for post-walked. 3360 // 3361 // During my forward walk I do a short 1-layer lookahead to see if I can find 3362 // a loop backedge with that doesn't have any work on the backedge. This 3363 // helps me construct nested loops with shared headers better. 3364 // 3365 // Once I've done the forward recursion, I do the post-work. For each child 3366 // I check to see if there is a backedge. Backedges define a loop! I 3367 // insert an IdealLoopTree at the target of the backedge. 3368 // 3369 // During the post-work I also check to see if I have several children 3370 // belonging to different loops. If so, then this Node is a decision point 3371 // where control flow can choose to change loop nests. It is at this 3372 // decision point where I can figure out how loops are nested. At this 3373 // time I can properly order the different loop nests from my children. 3374 // Note that there may not be any backedges at the decision point! 3375 // 3376 // Since the decision point can be far removed from the backedges, I can't 3377 // order my loops at the time I discover them. Thus at the decision point 3378 // I need to inspect loop header pre-order numbers to properly nest my 3379 // loops. This means I need to sort my childrens' loops by pre-order. 3380 // The sort is of size number-of-control-children, which generally limits 3381 // it to size 2 (i.e., I just choose between my 2 target loops). 3382 void PhaseIdealLoop::build_loop_tree() { 3383 // Allocate stack of size C->live_nodes()/2 to avoid frequent realloc 3384 GrowableArray <Node *> bltstack(C->live_nodes() >> 1); 3385 Node *n = C->root(); 3386 bltstack.push(n); 3387 int pre_order = 1; 3388 int stack_size; 3389 3390 while ( ( stack_size = bltstack.length() ) != 0 ) { 3391 n = bltstack.top(); // Leave node on stack 3392 if ( !is_visited(n) ) { 3393 // ---- Pre-pass Work ---- 3394 // Pre-walked but not post-walked nodes need a pre_order number. 3395 3396 set_preorder_visited( n, pre_order ); // set as visited 3397 3398 // ---- Scan over children ---- 3399 // Scan first over control projections that lead to loop headers. 3400 // This helps us find inner-to-outer loops with shared headers better. 3401 3402 // Scan children's children for loop headers. 3403 for ( int i = n->outcnt() - 1; i >= 0; --i ) { 3404 Node* m = n->raw_out(i); // Child 3405 if( m->is_CFG() && !is_visited(m) ) { // Only for CFG children 3406 // Scan over children's children to find loop 3407 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3408 Node* l = m->fast_out(j); 3409 if( is_visited(l) && // Been visited? 3410 !is_postvisited(l) && // But not post-visited 3411 get_preorder(l) < pre_order ) { // And smaller pre-order 3412 // Found! Scan the DFS down this path before doing other paths 3413 bltstack.push(m); 3414 break; 3415 } 3416 } 3417 } 3418 } 3419 pre_order++; 3420 } 3421 else if ( !is_postvisited(n) ) { 3422 // Note: build_loop_tree_impl() adds out edges on rare occasions, 3423 // such as com.sun.rsasign.am::a. 3424 // For non-recursive version, first, process current children. 3425 // On next iteration, check if additional children were added. 3426 for ( int k = n->outcnt() - 1; k >= 0; --k ) { 3427 Node* u = n->raw_out(k); 3428 if ( u->is_CFG() && !is_visited(u) ) { 3429 bltstack.push(u); 3430 } 3431 } 3432 if ( bltstack.length() == stack_size ) { 3433 // There were no additional children, post visit node now 3434 (void)bltstack.pop(); // Remove node from stack 3435 pre_order = build_loop_tree_impl( n, pre_order ); 3436 // Check for bailout 3437 if (C->failing()) { 3438 return; 3439 } 3440 // Check to grow _preorders[] array for the case when 3441 // build_loop_tree_impl() adds new nodes. 3442 check_grow_preorders(); 3443 } 3444 } 3445 else { 3446 (void)bltstack.pop(); // Remove post-visited node from stack 3447 } 3448 } 3449 } 3450 3451 //------------------------------build_loop_tree_impl--------------------------- 3452 int PhaseIdealLoop::build_loop_tree_impl( Node *n, int pre_order ) { 3453 // ---- Post-pass Work ---- 3454 // Pre-walked but not post-walked nodes need a pre_order number. 3455 3456 // Tightest enclosing loop for this Node 3457 IdealLoopTree *innermost = NULL; 3458 3459 // For all children, see if any edge is a backedge. If so, make a loop 3460 // for it. Then find the tightest enclosing loop for the self Node. 3461 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax; i++) { 3462 Node* m = n->fast_out(i); // Child 3463 if( n == m ) continue; // Ignore control self-cycles 3464 if( !m->is_CFG() ) continue;// Ignore non-CFG edges 3465 3466 IdealLoopTree *l; // Child's loop 3467 if( !is_postvisited(m) ) { // Child visited but not post-visited? 3468 // Found a backedge 3469 assert( get_preorder(m) < pre_order, "should be backedge" ); 3470 // Check for the RootNode, which is already a LoopNode and is allowed 3471 // to have multiple "backedges". 3472 if( m == C->root()) { // Found the root? 3473 l = _ltree_root; // Root is the outermost LoopNode 3474 } else { // Else found a nested loop 3475 // Insert a LoopNode to mark this loop. 3476 l = new IdealLoopTree(this, m, n); 3477 } // End of Else found a nested loop 3478 if( !has_loop(m) ) // If 'm' does not already have a loop set 3479 set_loop(m, l); // Set loop header to loop now 3480 3481 } else { // Else not a nested loop 3482 if( !_nodes[m->_idx] ) continue; // Dead code has no loop 3483 l = get_loop(m); // Get previously determined loop 3484 // If successor is header of a loop (nest), move up-loop till it 3485 // is a member of some outer enclosing loop. Since there are no 3486 // shared headers (I've split them already) I only need to go up 3487 // at most 1 level. 3488 while( l && l->_head == m ) // Successor heads loop? 3489 l = l->_parent; // Move up 1 for me 3490 // If this loop is not properly parented, then this loop 3491 // has no exit path out, i.e. its an infinite loop. 3492 if( !l ) { 3493 // Make loop "reachable" from root so the CFG is reachable. Basically 3494 // insert a bogus loop exit that is never taken. 'm', the loop head, 3495 // points to 'n', one (of possibly many) fall-in paths. There may be 3496 // many backedges as well. 3497 3498 // Here I set the loop to be the root loop. I could have, after 3499 // inserting a bogus loop exit, restarted the recursion and found my 3500 // new loop exit. This would make the infinite loop a first-class 3501 // loop and it would then get properly optimized. What's the use of 3502 // optimizing an infinite loop? 3503 l = _ltree_root; // Oops, found infinite loop 3504 3505 if (!_verify_only) { 3506 // Insert the NeverBranch between 'm' and it's control user. 3507 NeverBranchNode *iff = new NeverBranchNode( m ); 3508 _igvn.register_new_node_with_optimizer(iff); 3509 set_loop(iff, l); 3510 Node *if_t = new CProjNode( iff, 0 ); 3511 _igvn.register_new_node_with_optimizer(if_t); 3512 set_loop(if_t, l); 3513 3514 Node* cfg = NULL; // Find the One True Control User of m 3515 for (DUIterator_Fast jmax, j = m->fast_outs(jmax); j < jmax; j++) { 3516 Node* x = m->fast_out(j); 3517 if (x->is_CFG() && x != m && x != iff) 3518 { cfg = x; break; } 3519 } 3520 assert(cfg != NULL, "must find the control user of m"); 3521 uint k = 0; // Probably cfg->in(0) 3522 while( cfg->in(k) != m ) k++; // But check incase cfg is a Region 3523 cfg->set_req( k, if_t ); // Now point to NeverBranch 3524 _igvn._worklist.push(cfg); 3525 3526 // Now create the never-taken loop exit 3527 Node *if_f = new CProjNode( iff, 1 ); 3528 _igvn.register_new_node_with_optimizer(if_f); 3529 set_loop(if_f, l); 3530 // Find frame ptr for Halt. Relies on the optimizer 3531 // V-N'ing. Easier and quicker than searching through 3532 // the program structure. 3533 Node *frame = new ParmNode( C->start(), TypeFunc::FramePtr ); 3534 _igvn.register_new_node_with_optimizer(frame); 3535 // Halt & Catch Fire 3536 Node *halt = new HaltNode( if_f, frame ); 3537 _igvn.register_new_node_with_optimizer(halt); 3538 set_loop(halt, l); 3539 C->root()->add_req(halt); 3540 } 3541 set_loop(C->root(), _ltree_root); 3542 } 3543 } 3544 // Weeny check for irreducible. This child was already visited (this 3545 // IS the post-work phase). Is this child's loop header post-visited 3546 // as well? If so, then I found another entry into the loop. 3547 if (!_verify_only) { 3548 while( is_postvisited(l->_head) ) { 3549 // found irreducible 3550 l->_irreducible = 1; // = true 3551 l = l->_parent; 3552 _has_irreducible_loops = true; 3553 // Check for bad CFG here to prevent crash, and bailout of compile 3554 if (l == NULL) { 3555 C->record_method_not_compilable("unhandled CFG detected during loop optimization"); 3556 return pre_order; 3557 } 3558 } 3559 C->set_has_irreducible_loop(_has_irreducible_loops); 3560 } 3561 3562 // This Node might be a decision point for loops. It is only if 3563 // it's children belong to several different loops. The sort call 3564 // does a trivial amount of work if there is only 1 child or all 3565 // children belong to the same loop. If however, the children 3566 // belong to different loops, the sort call will properly set the 3567 // _parent pointers to show how the loops nest. 3568 // 3569 // In any case, it returns the tightest enclosing loop. 3570 innermost = sort( l, innermost ); 3571 } 3572 3573 // Def-use info will have some dead stuff; dead stuff will have no 3574 // loop decided on. 3575 3576 // Am I a loop header? If so fix up my parent's child and next ptrs. 3577 if( innermost && innermost->_head == n ) { 3578 assert( get_loop(n) == innermost, "" ); 3579 IdealLoopTree *p = innermost->_parent; 3580 IdealLoopTree *l = innermost; 3581 while( p && l->_head == n ) { 3582 l->_next = p->_child; // Put self on parents 'next child' 3583 p->_child = l; // Make self as first child of parent 3584 l = p; // Now walk up the parent chain 3585 p = l->_parent; 3586 } 3587 } else { 3588 // Note that it is possible for a LoopNode to reach here, if the 3589 // backedge has been made unreachable (hence the LoopNode no longer 3590 // denotes a Loop, and will eventually be removed). 3591 3592 // Record tightest enclosing loop for self. Mark as post-visited. 3593 set_loop(n, innermost); 3594 // Also record has_call flag early on 3595 if( innermost ) { 3596 if( n->is_Call() && !n->is_CallLeaf() && !n->is_macro() ) { 3597 // Do not count uncommon calls 3598 if( !n->is_CallStaticJava() || !n->as_CallStaticJava()->_name ) { 3599 Node *iff = n->in(0)->in(0); 3600 // No any calls for vectorized loops. 3601 if( UseSuperWord || !iff->is_If() || 3602 (n->in(0)->Opcode() == Op_IfFalse && 3603 (1.0 - iff->as_If()->_prob) >= 0.01) || 3604 (iff->as_If()->_prob >= 0.01) ) 3605 innermost->_has_call = 1; 3606 } 3607 } else if( n->is_Allocate() && n->as_Allocate()->_is_scalar_replaceable ) { 3608 // Disable loop optimizations if the loop has a scalar replaceable 3609 // allocation. This disabling may cause a potential performance lost 3610 // if the allocation is not eliminated for some reason. 3611 innermost->_allow_optimizations = false; 3612 innermost->_has_call = 1; // = true 3613 } else if (n->Opcode() == Op_SafePoint) { 3614 // Record all safepoints in this loop. 3615 if (innermost->_safepts == NULL) innermost->_safepts = new Node_List(); 3616 innermost->_safepts->push(n); 3617 } 3618 } 3619 } 3620 3621 // Flag as post-visited now 3622 set_postvisited(n); 3623 return pre_order; 3624 } 3625 3626 3627 //------------------------------build_loop_early------------------------------- 3628 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 3629 // First pass computes the earliest controlling node possible. This is the 3630 // controlling input with the deepest dominating depth. 3631 void PhaseIdealLoop::build_loop_early( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 3632 while (worklist.size() != 0) { 3633 // Use local variables nstack_top_n & nstack_top_i to cache values 3634 // on nstack's top. 3635 Node *nstack_top_n = worklist.pop(); 3636 uint nstack_top_i = 0; 3637 //while_nstack_nonempty: 3638 while (true) { 3639 // Get parent node and next input's index from stack's top. 3640 Node *n = nstack_top_n; 3641 uint i = nstack_top_i; 3642 uint cnt = n->req(); // Count of inputs 3643 if (i == 0) { // Pre-process the node. 3644 if( has_node(n) && // Have either loop or control already? 3645 !has_ctrl(n) ) { // Have loop picked out already? 3646 // During "merge_many_backedges" we fold up several nested loops 3647 // into a single loop. This makes the members of the original 3648 // loop bodies pointing to dead loops; they need to move up 3649 // to the new UNION'd larger loop. I set the _head field of these 3650 // dead loops to NULL and the _parent field points to the owning 3651 // loop. Shades of UNION-FIND algorithm. 3652 IdealLoopTree *ilt; 3653 while( !(ilt = get_loop(n))->_head ) { 3654 // Normally I would use a set_loop here. But in this one special 3655 // case, it is legal (and expected) to change what loop a Node 3656 // belongs to. 3657 _nodes.map(n->_idx, (Node*)(ilt->_parent) ); 3658 } 3659 // Remove safepoints ONLY if I've already seen I don't need one. 3660 // (the old code here would yank a 2nd safepoint after seeing a 3661 // first one, even though the 1st did not dominate in the loop body 3662 // and thus could be avoided indefinitely) 3663 if( !_verify_only && !_verify_me && ilt->_has_sfpt && n->Opcode() == Op_SafePoint && 3664 is_deleteable_safept(n)) { 3665 Node *in = n->in(TypeFunc::Control); 3666 lazy_replace(n,in); // Pull safepoint now 3667 if (ilt->_safepts != NULL) { 3668 ilt->_safepts->yank(n); 3669 } 3670 // Carry on with the recursion "as if" we are walking 3671 // only the control input 3672 if( !visited.test_set( in->_idx ) ) { 3673 worklist.push(in); // Visit this guy later, using worklist 3674 } 3675 // Get next node from nstack: 3676 // - skip n's inputs processing by setting i > cnt; 3677 // - we also will not call set_early_ctrl(n) since 3678 // has_node(n) == true (see the condition above). 3679 i = cnt + 1; 3680 } 3681 } 3682 } // if (i == 0) 3683 3684 // Visit all inputs 3685 bool done = true; // Assume all n's inputs will be processed 3686 while (i < cnt) { 3687 Node *in = n->in(i); 3688 ++i; 3689 if (in == NULL) continue; 3690 if (in->pinned() && !in->is_CFG()) 3691 set_ctrl(in, in->in(0)); 3692 int is_visited = visited.test_set( in->_idx ); 3693 if (!has_node(in)) { // No controlling input yet? 3694 assert( !in->is_CFG(), "CFG Node with no controlling input?" ); 3695 assert( !is_visited, "visit only once" ); 3696 nstack.push(n, i); // Save parent node and next input's index. 3697 nstack_top_n = in; // Process current input now. 3698 nstack_top_i = 0; 3699 done = false; // Not all n's inputs processed. 3700 break; // continue while_nstack_nonempty; 3701 } else if (!is_visited) { 3702 // This guy has a location picked out for him, but has not yet 3703 // been visited. Happens to all CFG nodes, for instance. 3704 // Visit him using the worklist instead of recursion, to break 3705 // cycles. Since he has a location already we do not need to 3706 // find his location before proceeding with the current Node. 3707 worklist.push(in); // Visit this guy later, using worklist 3708 } 3709 } 3710 if (done) { 3711 // All of n's inputs have been processed, complete post-processing. 3712 3713 // Compute earliest point this Node can go. 3714 // CFG, Phi, pinned nodes already know their controlling input. 3715 if (!has_node(n)) { 3716 // Record earliest legal location 3717 set_early_ctrl( n ); 3718 } 3719 if (nstack.is_empty()) { 3720 // Finished all nodes on stack. 3721 // Process next node on the worklist. 3722 break; 3723 } 3724 // Get saved parent node and next input's index. 3725 nstack_top_n = nstack.node(); 3726 nstack_top_i = nstack.index(); 3727 nstack.pop(); 3728 } 3729 } // while (true) 3730 } 3731 } 3732 3733 //------------------------------dom_lca_internal-------------------------------- 3734 // Pair-wise LCA 3735 Node *PhaseIdealLoop::dom_lca_internal( Node *n1, Node *n2 ) const { 3736 if( !n1 ) return n2; // Handle NULL original LCA 3737 assert( n1->is_CFG(), "" ); 3738 assert( n2->is_CFG(), "" ); 3739 // find LCA of all uses 3740 uint d1 = dom_depth(n1); 3741 uint d2 = dom_depth(n2); 3742 while (n1 != n2) { 3743 if (d1 > d2) { 3744 n1 = idom(n1); 3745 d1 = dom_depth(n1); 3746 } else if (d1 < d2) { 3747 n2 = idom(n2); 3748 d2 = dom_depth(n2); 3749 } else { 3750 // Here d1 == d2. Due to edits of the dominator-tree, sections 3751 // of the tree might have the same depth. These sections have 3752 // to be searched more carefully. 3753 3754 // Scan up all the n1's with equal depth, looking for n2. 3755 Node *t1 = idom(n1); 3756 while (dom_depth(t1) == d1) { 3757 if (t1 == n2) return n2; 3758 t1 = idom(t1); 3759 } 3760 // Scan up all the n2's with equal depth, looking for n1. 3761 Node *t2 = idom(n2); 3762 while (dom_depth(t2) == d2) { 3763 if (t2 == n1) return n1; 3764 t2 = idom(t2); 3765 } 3766 // Move up to a new dominator-depth value as well as up the dom-tree. 3767 n1 = t1; 3768 n2 = t2; 3769 d1 = dom_depth(n1); 3770 d2 = dom_depth(n2); 3771 } 3772 } 3773 return n1; 3774 } 3775 3776 //------------------------------compute_idom----------------------------------- 3777 // Locally compute IDOM using dom_lca call. Correct only if the incoming 3778 // IDOMs are correct. 3779 Node *PhaseIdealLoop::compute_idom( Node *region ) const { 3780 assert( region->is_Region(), "" ); 3781 Node *LCA = NULL; 3782 for( uint i = 1; i < region->req(); i++ ) { 3783 if( region->in(i) != C->top() ) 3784 LCA = dom_lca( LCA, region->in(i) ); 3785 } 3786 return LCA; 3787 } 3788 3789 bool PhaseIdealLoop::verify_dominance(Node* n, Node* use, Node* LCA, Node* early) { 3790 bool had_error = false; 3791 #ifdef ASSERT 3792 if (early != C->root()) { 3793 // Make sure that there's a dominance path from LCA to early 3794 Node* d = LCA; 3795 while (d != early) { 3796 if (d == C->root()) { 3797 dump_bad_graph("Bad graph detected in compute_lca_of_uses", n, early, LCA); 3798 tty->print_cr("*** Use %d isn't dominated by def %d ***", use->_idx, n->_idx); 3799 had_error = true; 3800 break; 3801 } 3802 d = idom(d); 3803 } 3804 } 3805 #endif 3806 return had_error; 3807 } 3808 3809 3810 Node* PhaseIdealLoop::compute_lca_of_uses(Node* n, Node* early, bool verify) { 3811 // Compute LCA over list of uses 3812 bool had_error = false; 3813 Node *LCA = NULL; 3814 for (DUIterator_Fast imax, i = n->fast_outs(imax); i < imax && LCA != early; i++) { 3815 Node* c = n->fast_out(i); 3816 if (_nodes[c->_idx] == NULL) 3817 continue; // Skip the occasional dead node 3818 if( c->is_Phi() ) { // For Phis, we must land above on the path 3819 for( uint j=1; j<c->req(); j++ ) {// For all inputs 3820 if( c->in(j) == n ) { // Found matching input? 3821 Node *use = c->in(0)->in(j); 3822 if (_verify_only && use->is_top()) continue; 3823 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3824 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3825 } 3826 } 3827 } else { 3828 // For CFG data-users, use is in the block just prior 3829 Node *use = has_ctrl(c) ? get_ctrl(c) : c->in(0); 3830 LCA = dom_lca_for_get_late_ctrl( LCA, use, n ); 3831 if (verify) had_error = verify_dominance(n, use, LCA, early) || had_error; 3832 } 3833 } 3834 assert(!had_error, "bad dominance"); 3835 return LCA; 3836 } 3837 3838 // Check the shape of the graph at the loop entry. In some cases, 3839 // the shape of the graph does not match the shape outlined below. 3840 // That is caused by the Opaque1 node "protecting" the shape of 3841 // the graph being removed by, for example, the IGVN performed 3842 // in PhaseIdealLoop::build_and_optimize(). 3843 // 3844 // After the Opaque1 node has been removed, optimizations (e.g., split-if, 3845 // loop unswitching, and IGVN, or a combination of them) can freely change 3846 // the graph's shape. As a result, the graph shape outlined below cannot 3847 // be guaranteed anymore. 3848 bool PhaseIdealLoop::is_canonical_loop_entry(CountedLoopNode* cl) { 3849 if (!cl->is_main_loop() && !cl->is_post_loop()) { 3850 return false; 3851 } 3852 Node* ctrl = cl->skip_predicates(); 3853 3854 if (ctrl == NULL || (!ctrl->is_IfTrue() && !ctrl->is_IfFalse())) { 3855 return false; 3856 } 3857 Node* iffm = ctrl->in(0); 3858 if (iffm == NULL || !iffm->is_If()) { 3859 return false; 3860 } 3861 Node* bolzm = iffm->in(1); 3862 if (bolzm == NULL || !bolzm->is_Bool()) { 3863 return false; 3864 } 3865 Node* cmpzm = bolzm->in(1); 3866 if (cmpzm == NULL || !cmpzm->is_Cmp()) { 3867 return false; 3868 } 3869 // compares can get conditionally flipped 3870 bool found_opaque = false; 3871 for (uint i = 1; i < cmpzm->req(); i++) { 3872 Node* opnd = cmpzm->in(i); 3873 if (opnd && opnd->Opcode() == Op_Opaque1) { 3874 found_opaque = true; 3875 break; 3876 } 3877 } 3878 if (!found_opaque) { 3879 return false; 3880 } 3881 return true; 3882 } 3883 3884 //------------------------------get_late_ctrl---------------------------------- 3885 // Compute latest legal control. 3886 Node *PhaseIdealLoop::get_late_ctrl( Node *n, Node *early ) { 3887 assert(early != NULL, "early control should not be NULL"); 3888 3889 Node* LCA = compute_lca_of_uses(n, early); 3890 #ifdef ASSERT 3891 if (LCA == C->root() && LCA != early) { 3892 // def doesn't dominate uses so print some useful debugging output 3893 compute_lca_of_uses(n, early, true); 3894 } 3895 #endif 3896 3897 // if this is a load, check for anti-dependent stores 3898 // We use a conservative algorithm to identify potential interfering 3899 // instructions and for rescheduling the load. The users of the memory 3900 // input of this load are examined. Any use which is not a load and is 3901 // dominated by early is considered a potentially interfering store. 3902 // This can produce false positives. 3903 if (n->is_Load() && LCA != early) { 3904 Node_List worklist; 3905 3906 Node *mem = n->in(MemNode::Memory); 3907 for (DUIterator_Fast imax, i = mem->fast_outs(imax); i < imax; i++) { 3908 Node* s = mem->fast_out(i); 3909 worklist.push(s); 3910 } 3911 while(worklist.size() != 0 && LCA != early) { 3912 Node* s = worklist.pop(); 3913 if (s->is_Load() || s->Opcode() == Op_SafePoint) { 3914 continue; 3915 } else if (s->is_MergeMem()) { 3916 for (DUIterator_Fast imax, i = s->fast_outs(imax); i < imax; i++) { 3917 Node* s1 = s->fast_out(i); 3918 worklist.push(s1); 3919 } 3920 } else { 3921 Node *sctrl = has_ctrl(s) ? get_ctrl(s) : s->in(0); 3922 assert(sctrl != NULL || s->outcnt() == 0, "must have control"); 3923 if (sctrl != NULL && !sctrl->is_top() && is_dominator(early, sctrl)) { 3924 LCA = dom_lca_for_get_late_ctrl(LCA, sctrl, n); 3925 } 3926 } 3927 } 3928 } 3929 3930 assert(LCA == find_non_split_ctrl(LCA), "unexpected late control"); 3931 return LCA; 3932 } 3933 3934 // true if CFG node d dominates CFG node n 3935 bool PhaseIdealLoop::is_dominator(Node *d, Node *n) { 3936 if (d == n) 3937 return true; 3938 assert(d->is_CFG() && n->is_CFG(), "must have CFG nodes"); 3939 uint dd = dom_depth(d); 3940 while (dom_depth(n) >= dd) { 3941 if (n == d) 3942 return true; 3943 n = idom(n); 3944 } 3945 return false; 3946 } 3947 3948 //------------------------------dom_lca_for_get_late_ctrl_internal------------- 3949 // Pair-wise LCA with tags. 3950 // Tag each index with the node 'tag' currently being processed 3951 // before advancing up the dominator chain using idom(). 3952 // Later calls that find a match to 'tag' know that this path has already 3953 // been considered in the current LCA (which is input 'n1' by convention). 3954 // Since get_late_ctrl() is only called once for each node, the tag array 3955 // does not need to be cleared between calls to get_late_ctrl(). 3956 // Algorithm trades a larger constant factor for better asymptotic behavior 3957 // 3958 Node *PhaseIdealLoop::dom_lca_for_get_late_ctrl_internal( Node *n1, Node *n2, Node *tag ) { 3959 uint d1 = dom_depth(n1); 3960 uint d2 = dom_depth(n2); 3961 3962 do { 3963 if (d1 > d2) { 3964 // current lca is deeper than n2 3965 _dom_lca_tags.map(n1->_idx, tag); 3966 n1 = idom(n1); 3967 d1 = dom_depth(n1); 3968 } else if (d1 < d2) { 3969 // n2 is deeper than current lca 3970 Node *memo = _dom_lca_tags[n2->_idx]; 3971 if( memo == tag ) { 3972 return n1; // Return the current LCA 3973 } 3974 _dom_lca_tags.map(n2->_idx, tag); 3975 n2 = idom(n2); 3976 d2 = dom_depth(n2); 3977 } else { 3978 // Here d1 == d2. Due to edits of the dominator-tree, sections 3979 // of the tree might have the same depth. These sections have 3980 // to be searched more carefully. 3981 3982 // Scan up all the n1's with equal depth, looking for n2. 3983 _dom_lca_tags.map(n1->_idx, tag); 3984 Node *t1 = idom(n1); 3985 while (dom_depth(t1) == d1) { 3986 if (t1 == n2) return n2; 3987 _dom_lca_tags.map(t1->_idx, tag); 3988 t1 = idom(t1); 3989 } 3990 // Scan up all the n2's with equal depth, looking for n1. 3991 _dom_lca_tags.map(n2->_idx, tag); 3992 Node *t2 = idom(n2); 3993 while (dom_depth(t2) == d2) { 3994 if (t2 == n1) return n1; 3995 _dom_lca_tags.map(t2->_idx, tag); 3996 t2 = idom(t2); 3997 } 3998 // Move up to a new dominator-depth value as well as up the dom-tree. 3999 n1 = t1; 4000 n2 = t2; 4001 d1 = dom_depth(n1); 4002 d2 = dom_depth(n2); 4003 } 4004 } while (n1 != n2); 4005 return n1; 4006 } 4007 4008 //------------------------------init_dom_lca_tags------------------------------ 4009 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 4010 // Intended use does not involve any growth for the array, so it could 4011 // be of fixed size. 4012 void PhaseIdealLoop::init_dom_lca_tags() { 4013 uint limit = C->unique() + 1; 4014 _dom_lca_tags.map( limit, NULL ); 4015 #ifdef ASSERT 4016 for( uint i = 0; i < limit; ++i ) { 4017 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 4018 } 4019 #endif // ASSERT 4020 } 4021 4022 //------------------------------clear_dom_lca_tags------------------------------ 4023 // Tag could be a node's integer index, 32bits instead of 64bits in some cases 4024 // Intended use does not involve any growth for the array, so it could 4025 // be of fixed size. 4026 void PhaseIdealLoop::clear_dom_lca_tags() { 4027 uint limit = C->unique() + 1; 4028 _dom_lca_tags.map( limit, NULL ); 4029 _dom_lca_tags.clear(); 4030 #ifdef ASSERT 4031 for( uint i = 0; i < limit; ++i ) { 4032 assert(_dom_lca_tags[i] == NULL, "Must be distinct from each node pointer"); 4033 } 4034 #endif // ASSERT 4035 } 4036 4037 //------------------------------build_loop_late-------------------------------- 4038 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4039 // Second pass finds latest legal placement, and ideal loop placement. 4040 void PhaseIdealLoop::build_loop_late( VectorSet &visited, Node_List &worklist, Node_Stack &nstack ) { 4041 while (worklist.size() != 0) { 4042 Node *n = worklist.pop(); 4043 // Only visit once 4044 if (visited.test_set(n->_idx)) continue; 4045 uint cnt = n->outcnt(); 4046 uint i = 0; 4047 while (true) { 4048 assert( _nodes[n->_idx], "no dead nodes" ); 4049 // Visit all children 4050 if (i < cnt) { 4051 Node* use = n->raw_out(i); 4052 ++i; 4053 // Check for dead uses. Aggressively prune such junk. It might be 4054 // dead in the global sense, but still have local uses so I cannot 4055 // easily call 'remove_dead_node'. 4056 if( _nodes[use->_idx] != NULL || use->is_top() ) { // Not dead? 4057 // Due to cycles, we might not hit the same fixed point in the verify 4058 // pass as we do in the regular pass. Instead, visit such phis as 4059 // simple uses of the loop head. 4060 if( use->in(0) && (use->is_CFG() || use->is_Phi()) ) { 4061 if( !visited.test(use->_idx) ) 4062 worklist.push(use); 4063 } else if( !visited.test_set(use->_idx) ) { 4064 nstack.push(n, i); // Save parent and next use's index. 4065 n = use; // Process all children of current use. 4066 cnt = use->outcnt(); 4067 i = 0; 4068 } 4069 } else { 4070 // Do not visit around the backedge of loops via data edges. 4071 // push dead code onto a worklist 4072 _deadlist.push(use); 4073 } 4074 } else { 4075 // All of n's children have been processed, complete post-processing. 4076 build_loop_late_post(n); 4077 if (nstack.is_empty()) { 4078 // Finished all nodes on stack. 4079 // Process next node on the worklist. 4080 break; 4081 } 4082 // Get saved parent node and next use's index. Visit the rest of uses. 4083 n = nstack.node(); 4084 cnt = n->outcnt(); 4085 i = nstack.index(); 4086 nstack.pop(); 4087 } 4088 } 4089 } 4090 } 4091 4092 // Verify that no data node is schedules in the outer loop of a strip 4093 // mined loop. 4094 void PhaseIdealLoop::verify_strip_mined_scheduling(Node *n, Node* least) { 4095 #ifdef ASSERT 4096 if (get_loop(least)->_nest == 0) { 4097 return; 4098 } 4099 IdealLoopTree* loop = get_loop(least); 4100 Node* head = loop->_head; 4101 if (head->is_OuterStripMinedLoop()) { 4102 Node* sfpt = head->as_Loop()->outer_safepoint(); 4103 ResourceMark rm; 4104 Unique_Node_List wq; 4105 wq.push(sfpt); 4106 for (uint i = 0; i < wq.size(); i++) { 4107 Node *m = wq.at(i); 4108 for (uint i = 1; i < m->req(); i++) { 4109 Node* nn = m->in(i); 4110 if (nn == n) { 4111 return; 4112 } 4113 if (nn != NULL && has_ctrl(nn) && get_loop(get_ctrl(nn)) == loop) { 4114 wq.push(nn); 4115 } 4116 } 4117 } 4118 ShouldNotReachHere(); 4119 } 4120 #endif 4121 } 4122 4123 4124 //------------------------------build_loop_late_post--------------------------- 4125 // Put Data nodes into some loop nest, by setting the _nodes[]->loop mapping. 4126 // Second pass finds latest legal placement, and ideal loop placement. 4127 void PhaseIdealLoop::build_loop_late_post( Node *n ) { 4128 4129 if (n->req() == 2 && (n->Opcode() == Op_ConvI2L || n->Opcode() == Op_CastII) && !C->major_progress() && !_verify_only) { 4130 _igvn._worklist.push(n); // Maybe we'll normalize it, if no more loops. 4131 } 4132 4133 #ifdef ASSERT 4134 if (_verify_only && !n->is_CFG()) { 4135 // Check def-use domination. 4136 compute_lca_of_uses(n, get_ctrl(n), true /* verify */); 4137 } 4138 #endif 4139 4140 // CFG and pinned nodes already handled 4141 if( n->in(0) ) { 4142 if( n->in(0)->is_top() ) return; // Dead? 4143 4144 // We'd like +VerifyLoopOptimizations to not believe that Mod's/Loads 4145 // _must_ be pinned (they have to observe their control edge of course). 4146 // Unlike Stores (which modify an unallocable resource, the memory 4147 // state), Mods/Loads can float around. So free them up. 4148 bool pinned = true; 4149 switch( n->Opcode() ) { 4150 case Op_DivI: 4151 case Op_DivF: 4152 case Op_DivD: 4153 case Op_ModI: 4154 case Op_ModF: 4155 case Op_ModD: 4156 case Op_LoadB: // Same with Loads; they can sink 4157 case Op_LoadUB: // during loop optimizations. 4158 case Op_LoadUS: 4159 case Op_LoadD: 4160 case Op_LoadF: 4161 case Op_LoadI: 4162 case Op_LoadKlass: 4163 case Op_LoadNKlass: 4164 case Op_LoadL: 4165 case Op_LoadS: 4166 case Op_LoadP: 4167 case Op_LoadBarrierSlowReg: 4168 case Op_LoadBarrierWeakSlowReg: 4169 case Op_LoadN: 4170 case Op_LoadRange: 4171 case Op_LoadD_unaligned: 4172 case Op_LoadL_unaligned: 4173 case Op_StrComp: // Does a bunch of load-like effects 4174 case Op_StrEquals: 4175 case Op_StrIndexOf: 4176 case Op_StrIndexOfChar: 4177 case Op_AryEq: 4178 case Op_HasNegatives: 4179 pinned = false; 4180 } 4181 if( pinned ) { 4182 IdealLoopTree *chosen_loop = get_loop(n->is_CFG() ? n : get_ctrl(n)); 4183 if( !chosen_loop->_child ) // Inner loop? 4184 chosen_loop->_body.push(n); // Collect inner loops 4185 return; 4186 } 4187 } else { // No slot zero 4188 if( n->is_CFG() ) { // CFG with no slot 0 is dead 4189 _nodes.map(n->_idx,0); // No block setting, it's globally dead 4190 return; 4191 } 4192 assert(!n->is_CFG() || n->outcnt() == 0, ""); 4193 } 4194 4195 // Do I have a "safe range" I can select over? 4196 Node *early = get_ctrl(n);// Early location already computed 4197 4198 // Compute latest point this Node can go 4199 Node *LCA = get_late_ctrl( n, early ); 4200 // LCA is NULL due to uses being dead 4201 if( LCA == NULL ) { 4202 #ifdef ASSERT 4203 for (DUIterator i1 = n->outs(); n->has_out(i1); i1++) { 4204 assert( _nodes[n->out(i1)->_idx] == NULL, "all uses must also be dead"); 4205 } 4206 #endif 4207 _nodes.map(n->_idx, 0); // This node is useless 4208 _deadlist.push(n); 4209 return; 4210 } 4211 assert(LCA != NULL && !LCA->is_top(), "no dead nodes"); 4212 4213 Node *legal = LCA; // Walk 'legal' up the IDOM chain 4214 Node *least = legal; // Best legal position so far 4215 while( early != legal ) { // While not at earliest legal 4216 #ifdef ASSERT 4217 if (legal->is_Start() && !early->is_Root()) { 4218 // Bad graph. Print idom path and fail. 4219 dump_bad_graph("Bad graph detected in build_loop_late", n, early, LCA); 4220 assert(false, "Bad graph detected in build_loop_late"); 4221 } 4222 #endif 4223 // Find least loop nesting depth 4224 legal = idom(legal); // Bump up the IDOM tree 4225 // Check for lower nesting depth 4226 if( get_loop(legal)->_nest < get_loop(least)->_nest ) 4227 least = legal; 4228 } 4229 assert(early == legal || legal != C->root(), "bad dominance of inputs"); 4230 4231 // Try not to place code on a loop entry projection 4232 // which can inhibit range check elimination. 4233 if (least != early) { 4234 Node* ctrl_out = least->unique_ctrl_out(); 4235 if (ctrl_out && ctrl_out->is_Loop() && 4236 least == ctrl_out->in(LoopNode::EntryControl)) { 4237 // Move the node above predicates as far up as possible so a 4238 // following pass of loop predication doesn't hoist a predicate 4239 // that depends on it above that node. 4240 Node* new_ctrl = least; 4241 for (;;) { 4242 if (!new_ctrl->is_Proj()) { 4243 break; 4244 } 4245 CallStaticJavaNode* call = new_ctrl->as_Proj()->is_uncommon_trap_if_pattern(Deoptimization::Reason_none); 4246 if (call == NULL) { 4247 break; 4248 } 4249 int req = call->uncommon_trap_request(); 4250 Deoptimization::DeoptReason trap_reason = Deoptimization::trap_request_reason(req); 4251 if (trap_reason != Deoptimization::Reason_loop_limit_check && 4252 trap_reason != Deoptimization::Reason_predicate && 4253 trap_reason != Deoptimization::Reason_profile_predicate) { 4254 break; 4255 } 4256 Node* c = new_ctrl->in(0)->in(0); 4257 if (is_dominator(c, early) && c != early) { 4258 break; 4259 } 4260 new_ctrl = c; 4261 } 4262 least = new_ctrl; 4263 } 4264 } 4265 4266 #ifdef ASSERT 4267 // If verifying, verify that 'verify_me' has a legal location 4268 // and choose it as our location. 4269 if( _verify_me ) { 4270 Node *v_ctrl = _verify_me->get_ctrl_no_update(n); 4271 Node *legal = LCA; 4272 while( early != legal ) { // While not at earliest legal 4273 if( legal == v_ctrl ) break; // Check for prior good location 4274 legal = idom(legal) ;// Bump up the IDOM tree 4275 } 4276 // Check for prior good location 4277 if( legal == v_ctrl ) least = legal; // Keep prior if found 4278 } 4279 #endif 4280 4281 // Assign discovered "here or above" point 4282 least = find_non_split_ctrl(least); 4283 verify_strip_mined_scheduling(n, least); 4284 set_ctrl(n, least); 4285 4286 // Collect inner loop bodies 4287 IdealLoopTree *chosen_loop = get_loop(least); 4288 if( !chosen_loop->_child ) // Inner loop? 4289 chosen_loop->_body.push(n);// Collect inner loops 4290 } 4291 4292 #ifdef ASSERT 4293 void PhaseIdealLoop::dump_bad_graph(const char* msg, Node* n, Node* early, Node* LCA) { 4294 tty->print_cr("%s", msg); 4295 tty->print("n: "); n->dump(); 4296 tty->print("early(n): "); early->dump(); 4297 if (n->in(0) != NULL && !n->in(0)->is_top() && 4298 n->in(0) != early && !n->in(0)->is_Root()) { 4299 tty->print("n->in(0): "); n->in(0)->dump(); 4300 } 4301 for (uint i = 1; i < n->req(); i++) { 4302 Node* in1 = n->in(i); 4303 if (in1 != NULL && in1 != n && !in1->is_top()) { 4304 tty->print("n->in(%d): ", i); in1->dump(); 4305 Node* in1_early = get_ctrl(in1); 4306 tty->print("early(n->in(%d)): ", i); in1_early->dump(); 4307 if (in1->in(0) != NULL && !in1->in(0)->is_top() && 4308 in1->in(0) != in1_early && !in1->in(0)->is_Root()) { 4309 tty->print("n->in(%d)->in(0): ", i); in1->in(0)->dump(); 4310 } 4311 for (uint j = 1; j < in1->req(); j++) { 4312 Node* in2 = in1->in(j); 4313 if (in2 != NULL && in2 != n && in2 != in1 && !in2->is_top()) { 4314 tty->print("n->in(%d)->in(%d): ", i, j); in2->dump(); 4315 Node* in2_early = get_ctrl(in2); 4316 tty->print("early(n->in(%d)->in(%d)): ", i, j); in2_early->dump(); 4317 if (in2->in(0) != NULL && !in2->in(0)->is_top() && 4318 in2->in(0) != in2_early && !in2->in(0)->is_Root()) { 4319 tty->print("n->in(%d)->in(%d)->in(0): ", i, j); in2->in(0)->dump(); 4320 } 4321 } 4322 } 4323 } 4324 } 4325 tty->cr(); 4326 tty->print("LCA(n): "); LCA->dump(); 4327 for (uint i = 0; i < n->outcnt(); i++) { 4328 Node* u1 = n->raw_out(i); 4329 if (u1 == n) 4330 continue; 4331 tty->print("n->out(%d): ", i); u1->dump(); 4332 if (u1->is_CFG()) { 4333 for (uint j = 0; j < u1->outcnt(); j++) { 4334 Node* u2 = u1->raw_out(j); 4335 if (u2 != u1 && u2 != n && u2->is_CFG()) { 4336 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4337 } 4338 } 4339 } else { 4340 Node* u1_later = get_ctrl(u1); 4341 tty->print("later(n->out(%d)): ", i); u1_later->dump(); 4342 if (u1->in(0) != NULL && !u1->in(0)->is_top() && 4343 u1->in(0) != u1_later && !u1->in(0)->is_Root()) { 4344 tty->print("n->out(%d)->in(0): ", i); u1->in(0)->dump(); 4345 } 4346 for (uint j = 0; j < u1->outcnt(); j++) { 4347 Node* u2 = u1->raw_out(j); 4348 if (u2 == n || u2 == u1) 4349 continue; 4350 tty->print("n->out(%d)->out(%d): ", i, j); u2->dump(); 4351 if (!u2->is_CFG()) { 4352 Node* u2_later = get_ctrl(u2); 4353 tty->print("later(n->out(%d)->out(%d)): ", i, j); u2_later->dump(); 4354 if (u2->in(0) != NULL && !u2->in(0)->is_top() && 4355 u2->in(0) != u2_later && !u2->in(0)->is_Root()) { 4356 tty->print("n->out(%d)->in(0): ", i); u2->in(0)->dump(); 4357 } 4358 } 4359 } 4360 } 4361 } 4362 tty->cr(); 4363 int ct = 0; 4364 Node *dbg_legal = LCA; 4365 while(!dbg_legal->is_Start() && ct < 100) { 4366 tty->print("idom[%d] ",ct); dbg_legal->dump(); 4367 ct++; 4368 dbg_legal = idom(dbg_legal); 4369 } 4370 tty->cr(); 4371 } 4372 #endif 4373 4374 #ifndef PRODUCT 4375 //------------------------------dump------------------------------------------- 4376 void PhaseIdealLoop::dump( ) const { 4377 ResourceMark rm; 4378 Arena* arena = Thread::current()->resource_area(); 4379 Node_Stack stack(arena, C->live_nodes() >> 2); 4380 Node_List rpo_list; 4381 VectorSet visited(arena); 4382 visited.set(C->top()->_idx); 4383 rpo( C->root(), stack, visited, rpo_list ); 4384 // Dump root loop indexed by last element in PO order 4385 dump( _ltree_root, rpo_list.size(), rpo_list ); 4386 } 4387 4388 void PhaseIdealLoop::dump( IdealLoopTree *loop, uint idx, Node_List &rpo_list ) const { 4389 loop->dump_head(); 4390 4391 // Now scan for CFG nodes in the same loop 4392 for( uint j=idx; j > 0; j-- ) { 4393 Node *n = rpo_list[j-1]; 4394 if( !_nodes[n->_idx] ) // Skip dead nodes 4395 continue; 4396 if( get_loop(n) != loop ) { // Wrong loop nest 4397 if( get_loop(n)->_head == n && // Found nested loop? 4398 get_loop(n)->_parent == loop ) 4399 dump(get_loop(n),rpo_list.size(),rpo_list); // Print it nested-ly 4400 continue; 4401 } 4402 4403 // Dump controlling node 4404 for( uint x = 0; x < loop->_nest; x++ ) 4405 tty->print(" "); 4406 tty->print("C"); 4407 if( n == C->root() ) { 4408 n->dump(); 4409 } else { 4410 Node* cached_idom = idom_no_update(n); 4411 Node *computed_idom = n->in(0); 4412 if( n->is_Region() ) { 4413 computed_idom = compute_idom(n); 4414 // computed_idom() will return n->in(0) when idom(n) is an IfNode (or 4415 // any MultiBranch ctrl node), so apply a similar transform to 4416 // the cached idom returned from idom_no_update. 4417 cached_idom = find_non_split_ctrl(cached_idom); 4418 } 4419 tty->print(" ID:%d",computed_idom->_idx); 4420 n->dump(); 4421 if( cached_idom != computed_idom ) { 4422 tty->print_cr("*** BROKEN IDOM! Computed as: %d, cached as: %d", 4423 computed_idom->_idx, cached_idom->_idx); 4424 } 4425 } 4426 // Dump nodes it controls 4427 for( uint k = 0; k < _nodes.Size(); k++ ) { 4428 // (k < C->unique() && get_ctrl(find(k)) == n) 4429 if (k < C->unique() && _nodes[k] == (Node*)((intptr_t)n + 1)) { 4430 Node *m = C->root()->find(k); 4431 if( m && m->outcnt() > 0 ) { 4432 if (!(has_ctrl(m) && get_ctrl_no_update(m) == n)) { 4433 tty->print_cr("*** BROKEN CTRL ACCESSOR! _nodes[k] is %p, ctrl is %p", 4434 _nodes[k], has_ctrl(m) ? get_ctrl_no_update(m) : NULL); 4435 } 4436 for( uint j = 0; j < loop->_nest; j++ ) 4437 tty->print(" "); 4438 tty->print(" "); 4439 m->dump(); 4440 } 4441 } 4442 } 4443 } 4444 } 4445 4446 // Collect a R-P-O for the whole CFG. 4447 // Result list is in post-order (scan backwards for RPO) 4448 void PhaseIdealLoop::rpo( Node *start, Node_Stack &stk, VectorSet &visited, Node_List &rpo_list ) const { 4449 stk.push(start, 0); 4450 visited.set(start->_idx); 4451 4452 while (stk.is_nonempty()) { 4453 Node* m = stk.node(); 4454 uint idx = stk.index(); 4455 if (idx < m->outcnt()) { 4456 stk.set_index(idx + 1); 4457 Node* n = m->raw_out(idx); 4458 if (n->is_CFG() && !visited.test_set(n->_idx)) { 4459 stk.push(n, 0); 4460 } 4461 } else { 4462 rpo_list.push(m); 4463 stk.pop(); 4464 } 4465 } 4466 } 4467 #endif 4468 4469 4470 //============================================================================= 4471 //------------------------------LoopTreeIterator----------------------------------- 4472 4473 // Advance to next loop tree using a preorder, left-to-right traversal. 4474 void LoopTreeIterator::next() { 4475 assert(!done(), "must not be done."); 4476 if (_curnt->_child != NULL) { 4477 _curnt = _curnt->_child; 4478 } else if (_curnt->_next != NULL) { 4479 _curnt = _curnt->_next; 4480 } else { 4481 while (_curnt != _root && _curnt->_next == NULL) { 4482 _curnt = _curnt->_parent; 4483 } 4484 if (_curnt == _root) { 4485 _curnt = NULL; 4486 assert(done(), "must be done."); 4487 } else { 4488 assert(_curnt->_next != NULL, "must be more to do"); 4489 _curnt = _curnt->_next; 4490 } 4491 } 4492 }